Been out of the loop on chips for a little while, can I get a quick refresher, this 10+nm process is still not using EUV correct? When is Intel expected/projected to release its first chips based off EUV?
First of all, Ice Lake started shipping 19Q4, so no, just over half a year. And secondly 10nm is only shipping in very limited capacity, and barely hit 4GHz, while 14++ can easily hit 5GHz+ on mobile. So NO it's still NOT ON TRACK.
Intel's 10nm is just there for Intel to say, look, I have it. That's all. It's by no means a viable process yet.
So jokes on you. Intel's 10nm is still just vaporware that can't even beat 14+, let alone 14++.
You rather mean 2022 - 2023 (they said "Q4 2022", which is code for a slip into the next year), which means at best 2H 2023, maybe Q4 2023 in high volume - which is the only one that matters - if they're lucky. If all stars and planets just shine right above Intel they might manage a Q3 2023 HVM.
Yeah no EUV. Hopefully 10+ (or technically 10++ if you include that one Cannonlake part) should allow higher clocks in comparison w Ice Lake where peak clocks took a bath compared to 14++++++.
7nm is due to come in Q421 with their discrete GPU line (I would imagine mid range not the biggest dies given the yield issues around a new process). This will be their first node w EUV.
I believe the GPUs that will use the 7 nm first are the massive data center GPUs that will be going into the Aurora supercomputer. They aren't trying to be cost-effective there, they are trying to flex their muscles. The graphics GPUs aren't likely to be put on 7 nm until later. I would imagine that low power CPUs would be put onto it before graphics GPUs.
Correct - the initial discrete GPUs will be 10nm/10nm+ and based off of TGL Xe. - Xe LP (integrated) and Xe HP.
The Ponte Vecchio GPU is the HPC variant, and will be the initial products on Intel's full EUV 7nm line. I would imagine that the 2nd iterations of the LP and HP will be moved to 7nm as well.
Sapphire Rapids (7nm Xeons) will probably be the 2nd product off the 7nm line, as the Sapphire Rapids is paired with Ponte Vecchio. Sapphire Rapids is PCIe5 + DDR5.
No EUV on 10nm. Full EUV (all layers, not just 1 or 2) on 7nm,
Discrete cards are coming soon - 10nm+ - Xe HP line - 4096 cores (512 EU), 8192 cores (1024 EU), 16,384 cores (2048 EU) - this is the huge chip Raj is holding up.
Xe HPC will be in Ponte Vecchio, which is 6 GPUs mesh linked, large unified cache (rambo) - will eventually see AICs based on that new design - but not 6 GPUs like in the Cray sled.
Is it me or it seems that Intel is churning new architectures every 2-3 months? Cascade lake, Comet Lake, Ice Lake were all released in 2019; three more (Cannon lake, Whiskey and Amber lake) in 2018. They are all anecdotally different: what's the point? Now Tiger lake is yeat another microArc, but finally on 10nm process: is it a shrink of any of the above?
Still, point is, they seem to have a different architecture for each product, vs AMD, for example, having the same exact silicon going into Ryzen, Threadrippers and Epyc. Yes, the connections are different, I/O chips are needed, but you get the point.
According to that link most of what you're talking about are grouped under the same architecture: cascade, comet, whiskey and amber are all Skylake. Icelake is also part of another group of 3 products with Icelake-SP and cooper. The only odd man out is cannon lake.
Cooper Lake is not Ice Lake - Ice Lake Xeons will be upto 38 cores and be limited to 1 and 2 socket designs, and are 10nm
Cooper Lake is still Skylake derived - 14nm - notable feature is bfloat16 and will be on the same platform as Ice Lake Xeon (Whitley) and use the same socket - but will be limited to 4 and 8 socket designs.
2 very different products for 2 very different markets.
Cannon lake was the first 10nm product - shipped a couple years ago - was not successful - it also had the Gen10 graphics (which is why we go from Gen9.5 on current Skylake designs to Gen 11 on Ice Lake - with no Gen10 apparent)
The last substantial and relevant desktop/laptop architecture Intel launched was Skylake. Intel is pushing for this weird naming confusion because it makes them look like they are actively launching newer and newer stuff, architectures, cores, etc. In reality they're pretty much the same CPUs as 5 years ago with slightly higher frequencies and performance that's been shot in both feet by the mitigations to their countless security vulnerabilities.
Pretty soon every model in the lineup will have one of these confusing names.
Ah, the irony of posting this comment on an article about a non-Skylake uArch CPU, that's a successor to an *existing* CPU, already with a non-Skylake uArch.
Sure, Willow Cove/TGL is what Intel now calls a new micro-archicture compared to Sunny Cove/ICL. Is that a reasonable description to people outside Intel? What's supposed to be new: - transistor optimizations (not micro-architecture OR architecture) - "security features" (so fixing more of the endless stream of security flaws in their baseline microarchitecture) - cache redesign
Now better cache is always nice. But it's not much to justify the claim of a new microarchitecture...
More than just an improved cache - wider and deeper than Skylake. So lack of knowledge clouds your understanding. Quite a new architecture, not just improved cache.
@Deicidium369 - TGL is indeed wider and deeper than Skylake, but the comparison @name99 made was to ICL, not Skylake. Is TGL substantially different from ICL? My understanding is that there were tweaks on the order of Haswell to Broadwell, rather than on the scale of Broadwell to Skylake.
@close seems to have been specific in saying "*substantial* and relevant desktop/laptop architecture". ICL hasn't been on the desktop and only serves part of the laptop market, so I think their comment was accurate in those terms even if a bit unfair overall.
Of course @mode_13h would rather bash on anyone appearing to criticise Intel, facts be damned... 🙄
WTF, dude? I mean, if you don't consider Sunny Cove a new uArch, then I guess you're going to call every future Intel x86 core just another Skylake iteration. But that's *your* problem, not Intel's.
According to your link they are all derivatives of Sandy Bridge.
From the link
"Like its predecessors, Sunny Cove focuses on extracting performance and reducing power through a number of key ways. Intel builds Sunny Cove on previous microarchitectures, descendants of Sandy Bridge. For the core to increase the overall performance, Intel focused on extracting additional parallelism. "
if they are all derivatives are sandy bridge, and even intel calls them 10th gen, how can they be a new architecture ? a new architecture would be going from Bulldozer-Excavator to Zen, correct ?
micro-architecture has a specific technical meaning, just like architecture has a specific technical meaning. If you want to treat it as a marketing term so that you can wave your tribal flag; if your goal is to score rather than to understand; go right ahead.
But be aware that the people who actually understand these issues will judge you accordingly. Respect, once lost, is hard to regain...
Comet Lake, for desktop computing - which is unreleased as of now, is still based on the 5-year old Skylake design. The OP is correct apart from a few mobile parts. Even Intel's latest mobile series have Skylake-class cores in them.
Zen2 is already superior to Skylake in essentially every metric, especially from an engineering/economics/efficiency perspective, while managing to be more performant per clock. By the time post-Skylake-class cores are fully deployed in all of Intel's product stacks; Zen3 will be shipping, and Zen4 on 5nm (which is in test/sampling right now) will be on the Horizon.
I want Intel to return to making decent products, but you can't argue the last 5 years of rehash have been decent products. Wake me up when they have the perf/watt, the perf/dollar and the features that AMD is offering - and not before.
Lisa Su / AMD have done more for client/.consumer computing in the last 3 years than Intel has done in the last Decade. It's probably because Dr. Su has is a computer science engineer, knows what she's talking about, and how to make a good processor - unlike Intel, which is run by greedy corporate bean-counters always seeking to rip the consumer (and OEMs) off as much as possible.
@AshlayW You're right about the last part, which I think it what chafes the fanboys most of all. We were stuck with quad-core on desktop and dual-core in low-power laptops for years until AMD announced Ryzen with 8 cores on desktop and 4 down the whole mobile stack, at which point Intel magically found a way to give us competitive products at similar prices. They got used to walking, and the moment they had to get back to running they tripped over their shoelaces.
Luckily for them the US rewards sheer size over and above all other metrics.
4th paragraph - LMAO - Lisa Su has a company based on making cheap chinese copies of what Intel (full of engineers) has made. She is better than her predecessor, so before the dumpster was full on blazing fire, and now the blaze is contained but still burning.
@Deicidium369 you really are a class idiot if thats what you believe, and i quote "Advanced Micro Devices was formally incorporated on May 1, 1969, by Jerry Sanders, along with seven of his colleagues from Fairchild Semiconductor.[3][4] Sanders, an electrical engineer who was the director of marketing at Fairchild, had, like many Fairchild executives, grown frustrated with the increasing lack of support, opportunity, and flexibility within the company, and decided to leave to start his own semiconductor company.[5] The previous year Robert Noyce, who had developed the first silicon integrated circuit in 1959 at Fairchild,[6] had left Fairchild together with Gordon Moore and founded the semiconductor company Intel in July 1968."
Both are USA corps both have had many firsts and at the moment AMDs CPUs are simply superior and Intel is playing catchup, AMD IS taking share from Intel a fact backed up by the Financial results of both sides, the OEMs are roped in to Intel right now with servers that are all linked, Intel is very clever at segmentation this ties servers to a upgrade or renew program and renew is more of a risk, however the tide is turning in corporate buying as it has in consumer, let us revisit this thread at year end and see the results
it's the triumph of marketing over engineering. They have nothing new to offer, so they flood the zone with PR month after month. The stream of new code names means nothing, as far as TECH goes the only interesting data points are:
- Lakefield (and only because it's new packaging; it will be an all-round sad package compared to say an A12Z or the A14) - Willow Cove (because it promises higher IPC -- let's see how much it actually delivers)
EVERYTHING else is old wine in new bottles. Sure, if you actually need to buy a new system, you may care about whether Rocket or Comet or Whiskey or whatever the latest name is in the segment you want to buy. But there's nothing new there in terms of interesting tech, just the same old same old slightly tweaked.
Nah, don't buy into his narrative. He's muddying the waters between the ticks and tocks.
Sure, everything is far more similar to the previous generation than not, but that's because CPUs are always converging towards what's optimal given the current technology. It doesn't make sense to start over with a clean-sheet design, because you'd probably just end up with something that looks similar to what you already have.
What? No, Intel wouldn't call it 1st gen, because it would require abandoning Core name and starting new marketing name for new architecture. As long as Core name has a positive ring for consumers, Intel will stick to this name. At some time Intel will switch to new name, but that's mostly marketing decision (to strengthen the new feeling of some uarch, for example Ocean Cove) .
The point is that *Intel* have muddied the waters between ticks and tocks - and they have, quite deliberately. Is TGL a tick or a tock? Meanwhile Skylake has gone tick/tick/tick/tick/tick all the way through to Comet Lake.
If ain't broken don't fix it. And Skylake/ice lake/tigerlake are all based on Sandy bridge more or less, which was a fantastic uarch. So good that and basically had to ditch its own invention, bulldozer and copy paste intel's Skylake core, pump it up with extra cache and call it a day.
OMG the Aliens have landed you have just admitted AMDs CPUs are competitive, open those eyes a little wider and you will see that actually they kick Intels ass
Calling Zen a copy/paste of Skylake with "extra cache" is one of the most reductive, ignorant and downright absurd statements I've seen on here. It makes no sense: not from a technical perspective (they look very different even from a schematic level), not from a practical perspective (how would they copy a design that was being produced at the same time they were working on Zen?) and not from a business perspective (they'd get sued).
Cool story, misinformed, an a little butthurt for some reason.
Ice Lake delivered 20-30% IPC and Tiger Lake likely to increase IPC by around the same amount
Rocket Lake S (8 core) is the Willow Coves from TGL and on the frequency optimized 14nm - so could launch at 4-4.5Ghz initially. This is the last 14nm desktop flagship. With the improved IPC from both Sunny and Willow Coves, will be a pretty big upgrade over the i9900K and Comet Lake 10 core.
Just re-read your last sentence, and you'll get the gist of my OP. We in 2-3 years we have a plethora of products with anecdotal performance increment. Ice Lake provided 2030% IPC improvement? Over what? Intl's own slides place it marginally above Whiskey lake and maybe 10% over Kaby Lake (https://www.anandtech.com/show/15385/intels-confus...
@yankeeDDL My understanding was that ICL had a 20-30% IPC increase married to a ~20% decrease in peak clock speeds vs Whiskey Lake, so overall a marginal performance increase on the CPU side and a significant one on the GPU side. In theory, *if* Rocket Lake S can take those increases and manage higher clocks on 14nm then it *could* be very competitive. But the die would also have to be much larger than even 10 core Comet Lake, either increasing cost or decreasing margins, and who knows what power consumption figures will look like. There's also no evidence yet that TGL will manage another 20-30% uplift either; early indications are more like 15%.
Overall I'm rating Deicidium's response 10/10 on the "strategic application of bullshit and projection" meter.
Well at much lower clocks my Ice Lake 1065G7 in a Dell 2-in-1 outperforms my 2 year old Dell 13 2-in-1 by a HUGE margin. I only buy the high end SKUs - so pretty much apples to apples.
Intel is coming out with architectures quickly now because they have a backlog of them from their 10 nm troubles. It's not every 2 to 3 months, but it's a fast clip. From the architecture introduced with skylake, which was several years back, they recently introduced sunny cove in ice lake, then tiger lake will have willow cove and golden cove will be introduced in 2021 I believe. After that they are rumored to have ocean cove planned for 2022-2023. So things will slow back down a bit.
Yep. They have pretty much 4 to 5 uarchitectures ready to go, just the node is keeping them still. What is new is the stuff Jim is working on and that is probably slated for 2022 or maybe 2023
Different variants / SKUs - New Microarchitecture is Ice Lake Sunny Cove and Tiger Lake Willow Cove. Pretty sure Keller was involved with Ice Lake after the reset
I feel as though Comet Lake was a contingency plan - that got far enough along that it made no sense to not release it. Not so much new architectures as new SKUs
So not only are they releasing new architectures Ice Lake and Tiger Lake, they're also releasing the contingency plan architectures - Comet Lake, Coffee Lake, etc..
Right - because the new ones can't cover half of their market segments, because their manufacturing capabilities are still compromised...
They're not releasing contingencies "because why not", they're releasing them because it's all they have to offer in those market segments and/or 10nm volume is insufficient to fulfil all orders.
Not new architectures - Cascade & Comet Lake is still based on Skylake.
Ice Lake is a new non Skylake architecture - Gen 1 of the new cores called Willow Cove. Ice Lake introduces a much enhanced IGP - Gen11. Tiger Lake will be shipping as a follow-up to Ice Lake with Gen 2 new cores called Willow Cove. Tiger Lake also introduced the Xe LP (Gen12) IGP.
There will be a 3rd enhancement to the all new core design - Golden Cove - not sure what product that will be in - assuming the follow up to Tiger Lake
Pcie 4.0 is not super important for 15w devices so I wouldn't count on it. It might support it but use a small number of lanes. Lpddr5 is important hopefully will be a feature.
It is becoming increasingly important for M.2 drives regardless of the CPU TDP, since they're limited to 4 lanes. More so when Sammy launches a new Evo with a new 4.0 controller that rocks the mainstream storage market again. So having at least a handful of 4.0 lanes would give these chips/chipsets/boards broader range and more longevity.
I'm not quite as sure LPDDR5 will be that important yet, although it depends on how much GPU they're willing to shove in a 15W chip. I'd wager LPDDR4X will still be sufficient, so I wouldn't be pissed if LPDDR5 doesn't make the cut.
As yeeeeman said, PCIe 4.0 is not relevant for 15 W laptops. It's too power-hungry. And if that weren't enough, the real-world benefit even for higher-powered devices is minimal.
The higher power variant of Tiger Lake for NUC11 will have a 4x PCIe4 M.2 storage as well as 2.5Gb ethernet. Have not seen mention of DDR5 of any flavor - likely first platform will be Sapphire Rapids which will be PCIe5 and DDR5. There is a small chance it could be on Rocket Lake.
>Finally, if everything goes according to plan or Intel, it looks like the Tiger Lake launch should be a higher volume affair than Ice Lake’s.
That is the key difference here, Icelake launch was really pushed just to show 10nm is actually OK. They were very hesitant to share out their shipping details. Being forthcoming about Tigerlake *shiping* to OEM mid year means they are quite confident about it.
That said we dont know what TDP series are they launching with Mobile.
AMD needs to move fast. Their marketing, and sales execution isn't doing good.
Shipped in late May early June to OEMs - bought the Dell 13 2-in-2 and got in October. But was also lower volume for 10nm, and 10nm+ is alot higher volume.
Tiger Lake will be in laptops/ultralights just like Ice Lake - but what I am looking forward to is the NUC11 with TGL - I have around 65 NUCs deployed over 3 generations - even the old ones are OK for our needs, but the graphics start to lag hard with dual 2560 or dual 4K - Gen12/Xe LP will be a welcome upgrade - the PCIe4 M.2 and 2.5Gb/s Ethernet is not bad either.
I think that Lisa Su's expectations of sales volume for Ryzen/Epyc is not what she was hoping for - it's always the next one that finally does whatever. Ryzen 1 - I dont think there was much expectations, but the refresh (2x00) I think was expected to be a sales hit - which it was not - same with Zen 2 - not huge sales successes - not failures - but no where near what she needed to start funding R&D. Ryzen 1 2 3 are just small iterations, and are what Keller left her with.
It's weird how the part of your post concerning Intel all makes sense, and then you just say... "stuff" about AMD. Zen 2 (3000 series) was AMD's Haswell moment - not a "small iteration". All indications point to a similar leap with Zen 3.
Overall sales volume compared with Intel isn't going to give you a useful idea of how they're doing, because AMD are playing from a position of historic weakness. Their relative growth reflects the strength of the products - Zen 2 has been a huge sales success on desktop. It'll take longer to make inroads into the server market due to the lengthy product cycles, but Zen 2 obliterates Intel there and it's not clear Intel have anything significant to respond with in the near future either.
AMD are definitely struggling with sales on the mobile side. Looks like poor alignment with OEM product cycles, lack of OEM trust after the mediocre 2000/3000 series, and lack of volume production are mostly to blame so far, as the product itself is sound.
Uh 10nm has been shipping in volume since Last May/June. the 10nm+ process is well ramped up and Tiger Lake / Ice Lake Xeons, Agilex FPGAs. etc - quite a lot of 10nm shipping and plenty of 10nm+ capacity.
You can call it whatever you want - Tiger Lake is Gen 11. Rocket Lake S - even though on 14nm - it is Tiger Lake under the hood - also Gen 11.
"In volume" is one of those terms that seems to be increasingly open to interpretation. It's certainly not at the volumes traditionally expected of Intel.
It hasn't shipped in high enough volumes to let Intel give up on 14nm respins even just in mobile. But there are still a pretty huge number of them out there, or you wouldn't be able to get Ice Lake chips in sub-$1K laptops now (rather than just expensive premium models like my XPS 2-in-1) and it wouldn't be in fairly high-volume laptops like Apple's newest.
than, not then. Intel's - the apostrophe (') indicated posession/ownership.
Still, (commas are your friend) better than Intel's 10nm availability. (a period denotes the end of a sentence).
Korguz still not adding anything at all to the discussion at hand. Don't worry, keep trying. I always root for someone with sub par intelligence when they try to add something to the conversation. Keep trying, little buddy! We are all rooting for you. You CAN Do It!
Meteor lake is slated for 2022 and 7nm. You wanted to say rocket lake which is a desktop chip based on the same core design as tigerlake but built on 14nm.
Know what's funny? TSMC had 16nm - made very tiny improvments and called it 14 and more small improvements and they call it 12.
So instead of + - TSMC makes up new names to show to their customers - since Intel only makes and sells Intel - the 14+ being 12nm and the ++ being 11nm wasn't needed. If you think that the latest iteration of Intel 14nm bears any resemblance to it's first iteration - you are mistaken.
You kiddies need to find a new meme - 10nm LOL is DEAD, and the 14++++++++++ was dead quite a while back. What else ya got?
This *isn't strictly correct*, though. Yes, TSMC rename processes that don't have substantial shrinks. Some of the changes Intel made to 14nm over time actually decrease density, though, and at least the first + was less of a physical process optimisation and more an optimisation of their ability to bin the resulting chips. You can fairly trivially undervolt an original quad-core mobile Skylake to run at Kaby Lake Refresh power levels.
You keep stating one or two faces and then spinning them into this bizarre pro-Intel narrative.
You need to get less personally upset about people making light of Intel's manufacturing woes. They're nearly over, but you're still *really* pressed about it.
I am not personally upset or actually even care one bit about any of this. Well, yes Intel HAD issues with 10nm - not any more. Kinda the inverse of TSMC having loser node after loser node and FINALLY getting on or 2 correct (16/14/12 and their 10nm class product "7nm"
I dont take any of this personally or care what you have to say. I wish you were a little bit more informed and could have a real discussion.
Do we know yet if Intel's Xe-LP graphics will continue to offer GPU virtualization to consumers? It's one of the secret upsides of an Intel iGPU and I'm concerned it will disappear as Intel gets more serious about its graphics (except for professional SKUs).
It's currently possible to use Broadwell or newer iGPUs in a virtualized setting with QEMU without *too* much of a fight, from what I've read. I'm planning on trying this configuration with a macOS virtual machine but haven't quite gotten to it.
Niche need for sure but if it ends up being supported, I'll be thrilled.
> with any luck, Intel’s 4+ years of playing with 10nm may finally pay some better dividends as they bring up their latest process.
It's funny you should use that analogy. If Intel hadn't been so intent on paying out dividends (and doing buybacks), their 10 nm might've come on sooner and stronger.
slide> * Committed to maintaining the dividend slide> * Repurchased $4.2B in shares
Well, it's good to see they've got their priorities straight. Uh...
They did spend a lot in buyback the last couple of years but it has nothing to do with their R&D spend. Their issues were not a result of not throwing enough money around. It was poor execution, and to fix it takes better execution and time, not just throwing around money.
As far as the buybacks and dividends, they have had huge profits. Giving back to investors is what they should be doing. They are already large with many pots on the fire. There comes a point where diversifying into more businesses will only dilute the company and won't provide the best value for shareholders. Giving returns to shareholders isn't corruption, it's sharing the profits of a successful business.
Yeah, I was sarcastic. I know it is a combination of many factors. I would add to what you said, bad management. They got used to great execution they even created the tick tock model. When 22nm ramped harder than previous nodes, they probably didn't adjust fab plans. 14nm was even worse but they still didn't learn their lesson because with 10nm they just went overboard with the features. any sane person would look at the past two nodes and say...let's learn something and do this one differently. Anyway...
100% with yeeeeman here. They let marketing set the roadmaps instead of engineering, and I get the impression there was an increasingly strong "yes culture" in middle-management that obfuscated the extent of the issues they were having in the foundry until it was too late, at which point they were deep into saving face at all costs.
Could be, but I think we're in the realm of pure speculation here. For all we know it was the engineering team leaders saying "push, push, we can do this."
The narrative about management woes has been corroborated by several long serving ex-Intel engineers - so while the usual cautions about trusting anecdotal evidence from former employees come into play, it does sound (and look) a lot like there were serious issues at the middle-management level that were at least significant contributing factors to their current difficulties.
The Intel complaints are pretty much all about how their managers suck. The Apple complaints (ever since Apple picked up the modem division) are ALSO pretty much all about how the Intel managers suck!
I did get a kick out of Apple buying their junk modem division. Apple will be taking revenge for that one soon enough when they kick Intel out of their low-end laptops.
All comes down to trying to make Gen 1 of 10nm 2.7X density increase instead of 2x. 10nm+ reaches that 2.7x density target. I also think that Cobalt was a lot more challenging - something that neither Samsung or TSMC has started on. So this time was used to benefit long term. Besides they could not make enough 14nm to fill all the orders, so 10nm would have been cool several years ago, but from a business point of view - un needed.
Depends on the other business units being expanded - 5G was sold to Apple, their Networking unit is very strong (even selling on most AMD systems) as are the Optane unit. So dilution is not much of an issue - this all plays into the taking control of more of the TAM for silicon.
"Besides they could not make enough 14nm to fill all the orders, so 10nm would have been cool several years ago, but from a business point of view - un needed."
How would having additional manufacturing capacity from a newer node - and being able to continue out-competing their closest rival on both transistor density and CPU core design - ever have been "un-needed"? 🤨
I know in the AMD world lithography is some magical panacea - but it's not.
You miss the point like a short outfielder misses a pop fly. Business only cares about revenues - and of your older much more optimized platform is selling in volumes you cannot keep up with - adding another newer products on the new process - would make little sense.]
From an enthusiast point of view - I would say that it WAS needed...
Intel's pretty notorious for making repeated and painful job cuts.
I don't have the data to back this up, but I think you should try to compare that R&D spend to a basket of companies that cover the equivalent business activities and markets to where Intel plays. I don't think your picture would look so rosy, then.
My sense is that each new manufacturing node is significantly more expensive than the last. And here's where Intel probably fell down. Why don't you plot TSMC's R&D spend, over the same time? If you can get Samsung's semiconductor manufacturing R&D, put that up too.
I never said that dividends were corruption, but I think you really can't argue that it's actually *good* for operations to do it, regardless of whether a company is executing to plan. When a company is having problems with execution, maybe the root cause isn't money, but often some extra spend can still help them get through a rough patch.
"My sense is that each new manufacturing node is significantly more expensive than the last."
A TSMC person, a few years ago, stated how many design rules needed to be used for the different nodes. The number of design rules (complexity of design) increased very rapidly with node shrinkage.
It's about time for Anandtech to post a detailed article about the actual node sizes involved in all of the different CPUs on the market and coming soon.
The fake nm naming stuff is intentionally confusing for consumers. It's not good enough to just relegate the truth to some random comments in articles.
Basically the revised 10nm node is just a bit denser than 14nm. Much much different than what they intended in the first place where they stated 100mtr/ mm2, better than tsmc 7nm.
in what way? power usage? Power usage is quite efficient on Ice Lake and assuming Tiger Lake.
and, no 10nm+ is not a bit denser than 14nm - 10+ reaches the initial target of 2.7x density increase.
Intel traditionally has multiple architectures - some frequency optimized (current 5GHz+) or power optimized (lakefield, Ice Lake) or density optimized. The 10nm is focused on power efficiency, 10nm+ will focus on density and (Ice Lake Servers) and frequency is likely never to reach 14nm levels - but then again neither can any node from TSMC
Your AMD numbers are wrong. Zen 2's *compute complex* is 52.7MTr/mm^2. You quoted the density of the *entire chip*, which includes the 14nm IO die. Renoir - which is entirely 7nm - is up at 63.33MTr/mm^2. https://www.reddit.com/r/Amd/comments/fraqll/amd_s...
As for your claims about Intel's 10nm+ density, I literally can't find any solid information that would confirm or refute it. You're effectively quoting their original 10nm marketing slides verbatim and expecting that to hold up in reality.
Intel's best case real world scenario 10nm density numbers are for Lakefield's compute die. This die omits all those large IO transistors, and has no requirements to be blazing fast. Even so, yeah, half the density of TSMC...
Cheers for the sauce. I'm amazed they missed their advertised numbers by that far, especially given the relatively small caches and offloaded IO. Given the purported numbers for Ice Lake, though, I do wonder if there's something missing from the story here.
Foveros looks quite promising as a future technology. Right now, though, it's amusing they had to do all of that just to reach comparable board space requirements (but not power/performance parity) with Apple's ARM designs.
Kirin 990 5G uses the 2nd generation 7nm process, N7+ (EUV). Lakefield uses 10nm+ as Intel's slide clearly shows.
Zen 2 chiplet is 52.7 MT/mm^2 using TSMC's first-gen 7nm. 14nm Xeons get just 15.8 MT/mm^2 (so about 3.3 times less dense).
We'll see how Tiger Lake does, but given it must use far more large and IO transistors than Lakefield, it'll be lucky to achieve 40 MT/mm^2, or about 2.5x scaling. Zen 3 will be out at the same time as Tiger Lake, if it uses 7nm+, density should improve to 60 MT/mm^2.
To be entirely fair, the Zen2 chiplet has very little external I/O and a LOT of SRAM cache (basically best case for density), while the 14nm Xeons (assuming LGA2066 or LGA3xxx, but even the client LGA11xx/1200 somewhat applies) have a lot of I/O, much of it that has to be driven over longer distances (which drives up feature size considerably). Intel chips also come with less cache vs their AMD counterparts. So 3.3x less dense isn't exactly a straight comparison.
Of course, AMD's design is made with large caches in mind, so it's not some sort of unfair comparison overall, since AMD's design uses that large cache to accommodate their slight latency penalty for having off-die I/O.
As for Tiger Lake, it has to accommodate a GPU, which are even more notoriously cache-light, lowering the density metric even further. Zen2 chiplet does not. The more interesting comparison, IMO, is Ice Lake & Tiger Lake density vs Zen2+Vega & Zen2/3+RDNA2. In these cases, AMD lowers their total cache by almost half, integrates a GPU, and also has to drive external I/O (DRAM, PCIe, USB, SATA, DP, eDP, etc). The comparison would have been less fair in the past, when the Intel U series chips had only DRAM, DP/eDP, and 4x PCIe lanes to account for (everything else was from the external chipset), but those are now joined by 4 TB3 controllers (each with 4 high speed differential pairs) basically not too dissimilar from having the SerDes for another 8 PCIe lanes, operating at a link rate between PCIe 4 and 5. Ice Lake also increases per core cache, too.
I think it's a fair comparison, as a result. Zen2 Renoir (APU) increased density to about 63.3MT/mm^2 (up from 52.7 MT/mm^2 on the Zen2 chiplet), this is inspite of decreased cache and increased I/O. Intel is treating Ice Lake-U transistor density like a trade secret, but they claim 7B transistors and measurements from Anandtech show 122.5mm^2 CPU size, giving ~57.1 MT/mm^2. Still not a straight comparison, as AMD doesn't seem to have made a low end, 7nm quad core APU this gen, and Intel only makes a quad core flagship 10nm product. Intel also still relies on a southbridge to drive the SATA, PCIe hub, USB, etc. It is within the ballpark for Intel's 10nm "high performance" variant's density claim (somewhere in the 60+ MT/mm^2 range, iirc).
Of course, this all has another problem: similar to "10nm" and "7nm" marketing names not meaning a whole lot, everyone measures "transistor count" differently, and even for the same company, the method has changed over time.
In the end, it basically means the density measurements we can get/make are almost meaningless.
I'm not sure the comparison between Kirin and Lakefield is necessarily that enlightening, given how much sustained effort ARM put into optimising their designs for area efficiency on a given process technology.
Renoir vs. Ice Lake ought to be a better point of comparison. Renoir runs 63.3MTr / mm^2 - I couldn't actually find numbers for ICL, though; just repetitions of Intel's original 100MTr / mm^2 claim which I believe was for NAND cells. I know it's lower than that for ICL on 10nm+, but have no idea by how much.
Lakefield is aiming to be a small low power chip like Kirin. It isn't possible to get a perfect comparison, but comparing mobile vs mobile and server vs server is quite reasonable. Either way, Renoir vs Lakefield doesn't look too good either. Despite Renoir being high-end and having more IO, it still has much better density.
Bottom line - the few 10nm numbers we have show it is well behind first-gen 7nm despite all the marketing claims of 100MT/mm^2.
That's a fair point - I just figure that what's effectively a first-gen product on new manufacturing tech (Foveros) is probably still not anywhere close to optimal. But you're right, that's the class of product it's effectively competing against...
Definitely agreed on the last point. 100MT/mm^2 was always an aspiration at best, and anyone repeating it seriously now needs their ears rinsed out.
So, it has now been FIVE YEARS since Intel released 14mn desktop silicon (Q2'15 65W and 95W) ... https://ark.intel.com/content/www/us/en/ark/produc... So when will Intel finally release desktop (or better, meaning at least 65W-95W parts) at any process node below 14nm? I am guessing Q2'22 or SEVEN YEARS after their 1st 14mn desktop parts.
If the first picture is the actual die size, I think it looks quite a lot bigger than the current Ice Lake U. Having used an Ice Lake U laptop, I feel that the 10nm seems to be closer to 14nm than 7nm. In fact, it seems to have regressed due to the amount of refinement over the years. Taking a quick comparison between Ice Lake and Whiskey Lake, the base clock have dropped significantly, and in some case, increased power requirement. In addition, the transition from 14nm to 10nm don't seem to help with the thermals. In my IdeaPad 5, I observed temps climbing up to 94 degs. While temp is a subjective metric highly dependent on the manufacturer's cooling solution and configuration, still 94 degs is the highest I have seen in a laptop. Even my older Ryzen 2500U don't go above 84 degs under sustained load.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
140 Comments
Back to Article
soliloquist - Friday, April 24, 2020 - link
Been out of the loop on chips for a little while, can I get a quick refresher, this 10+nm process is still not using EUV correct? When is Intel expected/projected to release its first chips based off EUV?Edkiefer - Friday, April 24, 2020 - link
`7nm, timeline 2021-22.shabby - Friday, April 24, 2020 - link
lolSpeedfriend - Saturday, April 25, 2020 - link
They have said that they are on track with 7nm and we have not seen any misses yet from ASML indicating that Intel aren't taking the EUV tools neededshabby - Saturday, April 25, 2020 - link
They've been "on track" with 10nm for the last 5 years, we all know how that turned out.Deicidium369 - Monday, May 4, 2020 - link
10nm started shipping a year ago, my Dell 14 2-in-1 has Ice Lake - Tiger Lake shipping to OEMs right now - Intel Agilex FPGA -10nm for over a year.So you need new material - that meme is dead.
dotjaz - Tuesday, May 5, 2020 - link
First of all, Ice Lake started shipping 19Q4, so no, just over half a year.And secondly 10nm is only shipping in very limited capacity, and barely hit 4GHz, while 14++ can easily hit 5GHz+ on mobile. So NO it's still NOT ON TRACK.
Intel's 10nm is just there for Intel to say, look, I have it. That's all. It's by no means a viable process yet.
So jokes on you. Intel's 10nm is still just vaporware that can't even beat 14+, let alone 14++.
Deicidium369 - Monday, May 4, 2020 - link
They hold a stake in ASML - can bet Intel gets preferential treatment regarding deliveries.Deicidium369 - Monday, May 4, 2020 - link
LOL all you want, kid, the shortages and issues are long behind them..Santoval - Sunday, April 26, 2020 - link
You rather mean 2022 - 2023 (they said "Q4 2022", which is code for a slip into the next year), which means at best 2H 2023, maybe Q4 2023 in high volume - which is the only one that matters - if they're lucky. If all stars and planets just shine right above Intel they might manage a Q3 2023 HVM.Deicidium369 - Monday, May 4, 2020 - link
LOL - nice mix of clueless and ignorant.Korguz - Monday, May 4, 2020 - link
you would know better then any one Deicidium369Jon Tseng - Friday, April 24, 2020 - link
Yeah no EUV. Hopefully 10+ (or technically 10++ if you include that one Cannonlake part) should allow higher clocks in comparison w Ice Lake where peak clocks took a bath compared to 14++++++.7nm is due to come in Q421 with their discrete GPU line (I would imagine mid range not the biggest dies given the yield issues around a new process). This will be their first node w EUV.
Yojimbo - Friday, April 24, 2020 - link
I believe the GPUs that will use the 7 nm first are the massive data center GPUs that will be going into the Aurora supercomputer. They aren't trying to be cost-effective there, they are trying to flex their muscles. The graphics GPUs aren't likely to be put on 7 nm until later. I would imagine that low power CPUs would be put onto it before graphics GPUs.Deicidium369 - Saturday, April 25, 2020 - link
Correct - the initial discrete GPUs will be 10nm/10nm+ and based off of TGL Xe. - Xe LP (integrated) and Xe HP.The Ponte Vecchio GPU is the HPC variant, and will be the initial products on Intel's full EUV 7nm line. I would imagine that the 2nd iterations of the LP and HP will be moved to 7nm as well.
Sapphire Rapids (7nm Xeons) will probably be the 2nd product off the 7nm line, as the Sapphire Rapids is paired with Ponte Vecchio. Sapphire Rapids is PCIe5 + DDR5.
Deicidium369 - Monday, May 4, 2020 - link
No EUV on 10nm. Full EUV (all layers, not just 1 or 2) on 7nm,Discrete cards are coming soon - 10nm+ - Xe HP line - 4096 cores (512 EU), 8192 cores (1024 EU), 16,384 cores (2048 EU) - this is the huge chip Raj is holding up.
Xe HPC will be in Ponte Vecchio, which is 6 GPUs mesh linked, large unified cache (rambo) - will eventually see AICs based on that new design - but not 6 GPUs like in the Cray sled.
yankeeDDL - Friday, April 24, 2020 - link
Is it me or it seems that Intel is churning new architectures every 2-3 months? Cascade lake, Comet Lake, Ice Lake were all released in 2019; three more (Cannon lake, Whiskey and Amber lake) in 2018. They are all anecdotally different: what's the point?Now Tiger lake is yeat another microArc, but finally on 10nm process: is it a shrink of any of the above?
IntelUser2000 - Friday, April 24, 2020 - link
You are confusing architectures with products.Cascade lake is server only.
Amberlake is 5W
Whiskeylake is 15W
Cometlake goes for high end laptops with 45W-plus TDP.
Amber/Whiskey/Comet are all identical microarchitectures. Cascade lake is server focused so bit different.
Icelake is the next gen, and Tigerlake is the follow up. Both on 10nm, latter a better version of it.
yankeeDDL - Friday, April 24, 2020 - link
According to Wikipedia, these are architectures:https://en.wikipedia.org/wiki/List_of_Intel_CPU_mi...
Still, point is, they seem to have a different architecture for each product, vs AMD, for example, having the same exact silicon going into Ryzen, Threadrippers and Epyc. Yes, the connections are different, I/O chips are needed, but you get the point.
whatthe123 - Friday, April 24, 2020 - link
According to that link most of what you're talking about are grouped under the same architecture:cascade, comet, whiskey and amber are all Skylake. Icelake is also part of another group of 3 products with Icelake-SP and cooper. The only odd man out is cannon lake.
Deicidium369 - Saturday, April 25, 2020 - link
Cooper Lake is not Ice Lake - Ice Lake Xeons will be upto 38 cores and be limited to 1 and 2 socket designs, and are 10nmCooper Lake is still Skylake derived - 14nm - notable feature is bfloat16 and will be on the same platform as Ice Lake Xeon (Whitley) and use the same socket - but will be limited to 4 and 8 socket designs.
2 very different products for 2 very different markets.
AshlayW - Monday, April 27, 2020 - link
Cooper / Ice Lake are overheating disasters that nobody wants.Deicidium369 - Monday, May 4, 2020 - link
LOL. They will outsell AMD 10:1 AT LEAST.Korguz - Monday, May 4, 2020 - link
oh ? and your source for this BS is what ?? post a link for once.Deicidium369 - Saturday, April 25, 2020 - link
Cannon lake was the first 10nm product - shipped a couple years ago - was not successful - it also had the Gen10 graphics (which is why we go from Gen9.5 on current Skylake designs to Gen 11 on Ice Lake - with no Gen10 apparent)close - Friday, April 24, 2020 - link
The last substantial and relevant desktop/laptop architecture Intel launched was Skylake. Intel is pushing for this weird naming confusion because it makes them look like they are actively launching newer and newer stuff, architectures, cores, etc. In reality they're pretty much the same CPUs as 5 years ago with slightly higher frequencies and performance that's been shot in both feet by the mitigations to their countless security vulnerabilities.Pretty soon every model in the lineup will have one of these confusing names.
mode_13h - Friday, April 24, 2020 - link
Ah, the irony of posting this comment on an article about a non-Skylake uArch CPU, that's a successor to an *existing* CPU, already with a non-Skylake uArch.name99 - Saturday, April 25, 2020 - link
Sure, Willow Cove/TGL is what Intel now calls a new micro-archicture compared to Sunny Cove/ICL. Is that a reasonable description to people outside Intel?What's supposed to be new:
- transistor optimizations (not micro-architecture OR architecture)
- "security features" (so fixing more of the endless stream of security flaws in their baseline microarchitecture)
- cache redesign
Now better cache is always nice. But it's not much to justify the claim of a new microarchitecture...
Deicidium369 - Saturday, April 25, 2020 - link
More than just an improved cache - wider and deeper than Skylake. So lack of knowledge clouds your understanding. Quite a new architecture, not just improved cache.mode_13h - Saturday, April 25, 2020 - link
I think it's more a case of *willful* ignorance. He just wants to bash on Intel, facts be damned.Deicidium369 - Monday, May 4, 2020 - link
Yeah, pretty obvious. AMD GOOD! INTEL BAD! NVIDIA BAD! AMD OUTSELL INTEL, but AMD just being modest and not really reporting the revenue ...Spunjji - Monday, April 27, 2020 - link
@Deicidium369 - TGL is indeed wider and deeper than Skylake, but the comparison @name99 made was to ICL, not Skylake. Is TGL substantially different from ICL? My understanding is that there were tweaks on the order of Haswell to Broadwell, rather than on the scale of Broadwell to Skylake.@close seems to have been specific in saying "*substantial* and relevant desktop/laptop architecture". ICL hasn't been on the desktop and only serves part of the laptop market, so I think their comment was accurate in those terms even if a bit unfair overall.
Of course @mode_13h would rather bash on anyone appearing to criticise Intel, facts be damned... 🙄
mode_13h - Saturday, April 25, 2020 - link
WTF, dude? I mean, if you don't consider Sunny Cove a new uArch, then I guess you're going to call every future Intel x86 core just another Skylake iteration. But that's *your* problem, not Intel's.https://en.wikichip.org/wiki/intel/microarchitectu...
29a - Sunday, April 26, 2020 - link
According to your link they are all derivatives of Sandy Bridge.From the link
"Like its predecessors, Sunny Cove focuses on extracting performance and reducing power through a number of key ways. Intel builds Sunny Cove on previous microarchitectures, descendants of Sandy Bridge. For the core to increase the overall performance, Intel focused on extracting additional parallelism. "
Qasar - Sunday, April 26, 2020 - link
if they are all derivatives are sandy bridge, and even intel calls them 10th gen, how can they be a new architecture ? a new architecture would be going from Bulldozer-Excavator to Zen, correct ?name99 - Monday, April 27, 2020 - link
micro-architecture has a specific technical meaning, just like architecture has a specific technical meaning.If you want to treat it as a marketing term so that you can wave your tribal flag; if your goal is to score rather than to understand; go right ahead.
But be aware that the people who actually understand these issues will judge you accordingly. Respect, once lost, is hard to regain...
Qasar - Tuesday, April 28, 2020 - link
name99, still doesnt explain why intel calls it 10th gen and not 1st gen, sounds like its still based on the previous architecture to me.AshlayW - Monday, April 27, 2020 - link
Comet Lake, for desktop computing - which is unreleased as of now, is still based on the 5-year old Skylake design. The OP is correct apart from a few mobile parts. Even Intel's latest mobile series have Skylake-class cores in them.Zen2 is already superior to Skylake in essentially every metric, especially from an engineering/economics/efficiency perspective, while managing to be more performant per clock. By the time post-Skylake-class cores are fully deployed in all of Intel's product stacks; Zen3 will be shipping, and Zen4 on 5nm (which is in test/sampling right now) will be on the Horizon.
I want Intel to return to making decent products, but you can't argue the last 5 years of rehash have been decent products. Wake me up when they have the perf/watt, the perf/dollar and the features that AMD is offering - and not before.
Lisa Su / AMD have done more for client/.consumer computing in the last 3 years than Intel has done in the last Decade. It's probably because Dr. Su has is a computer science engineer, knows what she's talking about, and how to make a good processor - unlike Intel, which is run by greedy corporate bean-counters always seeking to rip the consumer (and OEMs) off as much as possible.
Spunjji - Tuesday, April 28, 2020 - link
@AshlayW You're right about the last part, which I think it what chafes the fanboys most of all. We were stuck with quad-core on desktop and dual-core in low-power laptops for years until AMD announced Ryzen with 8 cores on desktop and 4 down the whole mobile stack, at which point Intel magically found a way to give us competitive products at similar prices. They got used to walking, and the moment they had to get back to running they tripped over their shoelaces.Luckily for them the US rewards sheer size over and above all other metrics.
Deicidium369 - Monday, May 4, 2020 - link
1st paragraph - corrent2nd paragraph - LMAO
4th paragraph - LMAO - Lisa Su has a company based on making cheap chinese copies of what Intel (full of engineers) has made. She is better than her predecessor, so before the dumpster was full on blazing fire, and now the blaze is contained but still burning.
alufan - Monday, May 4, 2020 - link
@Deicidium369 you really are a class idiot if thats what you believe, and i quote"Advanced Micro Devices was formally incorporated on May 1, 1969, by Jerry Sanders, along with seven of his colleagues from Fairchild Semiconductor.[3][4] Sanders, an electrical engineer who was the director of marketing at Fairchild, had, like many Fairchild executives, grown frustrated with the increasing lack of support, opportunity, and flexibility within the company, and decided to leave to start his own semiconductor company.[5] The previous year Robert Noyce, who had developed the first silicon integrated circuit in 1959 at Fairchild,[6] had left Fairchild together with Gordon Moore and founded the semiconductor company Intel in July 1968."
Both are USA corps both have had many firsts and at the moment AMDs CPUs are simply superior and Intel is playing catchup, AMD IS taking share from Intel a fact backed up by the Financial results of both sides, the OEMs are roped in to Intel right now with servers that are all linked, Intel is very clever at segmentation this ties servers to a upgrade or renew program and renew is more of a risk, however the tide is turning in corporate buying as it has in consumer, let us revisit this thread at year end and see the results
name99 - Saturday, April 25, 2020 - link
it's the triumph of marketing over engineering. They have nothing new to offer, so they flood the zone with PR month after month.The stream of new code names means nothing, as far as TECH goes the only interesting data points are:
- Lakefield (and only because it's new packaging; it will be an all-round sad package compared to say an A12Z or the A14)
- Willow Cove (because it promises higher IPC -- let's see how much it actually delivers)
EVERYTHING else is old wine in new bottles. Sure, if you actually need to buy a new system, you may care about whether Rocket or Comet or Whiskey or whatever the latest name is in the segment you want to buy. But there's nothing new there in terms of interesting tech, just the same old same old slightly tweaked.
Qasar - Saturday, April 25, 2020 - link
thats why its called 10th gen, not 1st gen.mode_13h - Saturday, April 25, 2020 - link
Nah, don't buy into his narrative. He's muddying the waters between the ticks and tocks.Sure, everything is far more similar to the previous generation than not, but that's because CPUs are always converging towards what's optimal given the current technology. It doesn't make sense to start over with a clean-sheet design, because you'd probably just end up with something that looks similar to what you already have.
Qasar - Saturday, April 25, 2020 - link
who said i was ? IF it was a NEW architecture, wouldnt intel of called it 1st gen instead of 10th gen ?Deicidium369 - Monday, May 4, 2020 - link
No you are on your own short bus narrativeTabalan - Tuesday, May 5, 2020 - link
What? No, Intel wouldn't call it 1st gen, because it would require abandoning Core name and starting new marketing name for new architecture. As long as Core name has a positive ring for consumers, Intel will stick to this name. At some time Intel will switch to new name, but that's mostly marketing decision (to strengthen the new feeling of some uarch, for example Ocean Cove) .Spunjji - Monday, April 27, 2020 - link
The point is that *Intel* have muddied the waters between ticks and tocks - and they have, quite deliberately. Is TGL a tick or a tock? Meanwhile Skylake has gone tick/tick/tick/tick/tick all the way through to Comet Lake.Deicidium369 - Monday, May 4, 2020 - link
That model was laid to rest quite a while ago.yeeeeman - Saturday, April 25, 2020 - link
If ain't broken don't fix it. And Skylake/ice lake/tigerlake are all based on Sandy bridge more or less, which was a fantastic uarch. So good that and basically had to ditch its own invention, bulldozer and copy paste intel's Skylake core, pump it up with extra cache and call it a day.yeeeeman - Saturday, April 25, 2020 - link
AMD had to ditch*Deicidium369 - Saturday, April 25, 2020 - link
Well when Core dropped, AMD's time in the sun came to and end for more than a decadeThud2 - Saturday, April 25, 2020 - link
Yeah I remember 2006. Nickleback, those were the days?Deicidium369 - Monday, May 4, 2020 - link
My ex loved Nickleback - threw so many of those CDs out the windown.2006 was the last time had a remotely competitive products, then 12 years before It has another one. 12 YEARS.
alufan - Monday, May 4, 2020 - link
OMG the Aliens have landed you have just admitted AMDs CPUs are competitive, open those eyes a little wider and you will see that actually they kick Intels assSpunjji - Monday, April 27, 2020 - link
Calling Zen a copy/paste of Skylake with "extra cache" is one of the most reductive, ignorant and downright absurd statements I've seen on here. It makes no sense: not from a technical perspective (they look very different even from a schematic level), not from a practical perspective (how would they copy a design that was being produced at the same time they were working on Zen?) and not from a business perspective (they'd get sued).Are you Gondalf in disguise?
Deicidium369 - Saturday, April 25, 2020 - link
Cool story, misinformed, an a little butthurt for some reason.Ice Lake delivered 20-30% IPC and Tiger Lake likely to increase IPC by around the same amount
Rocket Lake S (8 core) is the Willow Coves from TGL and on the frequency optimized 14nm - so could launch at 4-4.5Ghz initially. This is the last 14nm desktop flagship. With the improved IPC from both Sunny and Willow Coves, will be a pretty big upgrade over the i9900K and Comet Lake 10 core.
yankeeDDL - Sunday, April 26, 2020 - link
Just re-read your last sentence, and you'll get the gist of my OP.We in 2-3 years we have a plethora of products with anecdotal performance increment. Ice Lake provided 2030% IPC improvement? Over what? Intl's own slides place it marginally above Whiskey lake and maybe 10% over Kaby Lake (https://www.anandtech.com/show/15385/intels-confus...
Spunjji - Monday, April 27, 2020 - link
@yankeeDDL My understanding was that ICL had a 20-30% IPC increase married to a ~20% decrease in peak clock speeds vs Whiskey Lake, so overall a marginal performance increase on the CPU side and a significant one on the GPU side. In theory, *if* Rocket Lake S can take those increases and manage higher clocks on 14nm then it *could* be very competitive. But the die would also have to be much larger than even 10 core Comet Lake, either increasing cost or decreasing margins, and who knows what power consumption figures will look like. There's also no evidence yet that TGL will manage another 20-30% uplift either; early indications are more like 15%.Overall I'm rating Deicidium's response 10/10 on the "strategic application of bullshit and projection" meter.
Deicidium369 - Monday, May 4, 2020 - link
Oh no some Rando is rating my post...Deicidium369 - Monday, May 4, 2020 - link
Well at much lower clocks my Ice Lake 1065G7 in a Dell 2-in-1 outperforms my 2 year old Dell 13 2-in-1 by a HUGE margin. I only buy the high end SKUs - so pretty much apples to apples.Deicidium369 - Monday, May 4, 2020 - link
Sunny Cover delivered a 30% IPC increase - would suspect Tiger Lake Willow coves to offer similar gains.Yojimbo - Friday, April 24, 2020 - link
Intel is coming out with architectures quickly now because they have a backlog of them from their 10 nm troubles. It's not every 2 to 3 months, but it's a fast clip. From the architecture introduced with skylake, which was several years back, they recently introduced sunny cove in ice lake, then tiger lake will have willow cove and golden cove will be introduced in 2021 I believe. After that they are rumored to have ocean cove planned for 2022-2023. So things will slow back down a bit.yeeeeman - Saturday, April 25, 2020 - link
Yep. They have pretty much 4 to 5 uarchitectures ready to go, just the node is keeping them still.What is new is the stuff Jim is working on and that is probably slated for 2022 or maybe 2023
Deicidium369 - Saturday, April 25, 2020 - link
Different variants / SKUs - New Microarchitecture is Ice Lake Sunny Cove and Tiger Lake Willow Cove. Pretty sure Keller was involved with Ice Lake after the resetDeicidium369 - Saturday, April 25, 2020 - link
I feel as though Comet Lake was a contingency plan - that got far enough along that it made no sense to not release it. Not so much new architectures as new SKUsvFunct - Sunday, April 26, 2020 - link
So not only are they releasing new architectures Ice Lake and Tiger Lake, they're also releasing the contingency plan architectures - Comet Lake, Coffee Lake, etc..Spunjji - Tuesday, April 28, 2020 - link
Right - because the new ones can't cover half of their market segments, because their manufacturing capabilities are still compromised...They're not releasing contingencies "because why not", they're releasing them because it's all they have to offer in those market segments and/or 10nm volume is insufficient to fulfil all orders.
Deicidium369 - Saturday, April 25, 2020 - link
Not new architectures - Cascade & Comet Lake is still based on Skylake.Ice Lake is a new non Skylake architecture - Gen 1 of the new cores called Willow Cove. Ice Lake introduces a much enhanced IGP - Gen11.
Tiger Lake will be shipping as a follow-up to Ice Lake with Gen 2 new cores called Willow Cove. Tiger Lake also introduced the Xe LP (Gen12) IGP.
There will be a 3rd enhancement to the all new core design - Golden Cove - not sure what product that will be in - assuming the follow up to Tiger Lake
JayNor - Friday, April 24, 2020 - link
Intel says Tiger Lake has a new Willow Cove core, bigger caches, new Xe GPU. tweaked 10nm process.Rumors/leaks also say PCIE4, LPDDR5.
yeeeeman - Friday, April 24, 2020 - link
Pcie 4.0 is not super important for 15w devices so I wouldn't count on it. It might support it but use a small number of lanes. Lpddr5 is important hopefully will be a feature.Alexvrb - Friday, April 24, 2020 - link
It is becoming increasingly important for M.2 drives regardless of the CPU TDP, since they're limited to 4 lanes. More so when Sammy launches a new Evo with a new 4.0 controller that rocks the mainstream storage market again. So having at least a handful of 4.0 lanes would give these chips/chipsets/boards broader range and more longevity.I'm not quite as sure LPDDR5 will be that important yet, although it depends on how much GPU they're willing to shove in a 15W chip. I'd wager LPDDR4X will still be sufficient, so I wouldn't be pissed if LPDDR5 doesn't make the cut.
yeeeeman - Saturday, April 25, 2020 - link
Lpddr5 will bring efficiency inprovements. Bandwidth is more than enough with lpddr4x alreadymode_13h - Saturday, April 25, 2020 - link
As yeeeeman said, PCIe 4.0 is not relevant for 15 W laptops. It's too power-hungry. And if that weren't enough, the real-world benefit even for higher-powered devices is minimal.Deicidium369 - Saturday, April 25, 2020 - link
The higher power variant of Tiger Lake for NUC11 will have a 4x PCIe4 M.2 storage as well as 2.5Gb ethernet. Have not seen mention of DDR5 of any flavor - likely first platform will be Sapphire Rapids which will be PCIe5 and DDR5. There is a small chance it could be on Rocket Lake.Deicidium369 - Saturday, April 25, 2020 - link
Tiger Lake is PCIe4 but likely not LDPPR5. The follow up to TGL will likely have LPDDR5.Much more than just larger caches - both wider and deeper than Skylake.
ksec - Friday, April 24, 2020 - link
>Finally, if everything goes according to plan or Intel, it looks like the Tiger Lake launch should be a higher volume affair than Ice Lake’s.That is the key difference here, Icelake launch was really pushed just to show 10nm is actually OK. They were very hesitant to share out their shipping details. Being forthcoming about Tigerlake *shiping* to OEM mid year means they are quite confident about it.
That said we dont know what TDP series are they launching with Mobile.
AMD needs to move fast. Their marketing, and sales execution isn't doing good.
Deicidium369 - Saturday, April 25, 2020 - link
Shipped in late May early June to OEMs - bought the Dell 13 2-in-2 and got in October. But was also lower volume for 10nm, and 10nm+ is alot higher volume.Tiger Lake will be in laptops/ultralights just like Ice Lake - but what I am looking forward to is the NUC11 with TGL - I have around 65 NUCs deployed over 3 generations - even the old ones are OK for our needs, but the graphics start to lag hard with dual 2560 or dual 4K - Gen12/Xe LP will be a welcome upgrade - the PCIe4 M.2 and 2.5Gb/s Ethernet is not bad either.
I think that Lisa Su's expectations of sales volume for Ryzen/Epyc is not what she was hoping for - it's always the next one that finally does whatever. Ryzen 1 - I dont think there was much expectations, but the refresh (2x00) I think was expected to be a sales hit - which it was not - same with Zen 2 - not huge sales successes - not failures - but no where near what she needed to start funding R&D. Ryzen 1 2 3 are just small iterations, and are what Keller left her with.
Spunjji - Monday, April 27, 2020 - link
It's weird how the part of your post concerning Intel all makes sense, and then you just say... "stuff" about AMD. Zen 2 (3000 series) was AMD's Haswell moment - not a "small iteration". All indications point to a similar leap with Zen 3.Overall sales volume compared with Intel isn't going to give you a useful idea of how they're doing, because AMD are playing from a position of historic weakness. Their relative growth reflects the strength of the products - Zen 2 has been a huge sales success on desktop. It'll take longer to make inroads into the server market due to the lengthy product cycles, but Zen 2 obliterates Intel there and it's not clear Intel have anything significant to respond with in the near future either.
AMD are definitely struggling with sales on the mobile side. Looks like poor alignment with OEM product cycles, lack of OEM trust after the mediocre 2000/3000 series, and lack of volume production are mostly to blame so far, as the product itself is sound.
Deicidium369 - Monday, May 4, 2020 - link
AMD is struggling with sales on every single product they offer.Korguz - Monday, May 4, 2020 - link
Deicidium369 post proof of your BS personal pro intel opinionsTristanSDX - Friday, April 24, 2020 - link
lol, still accelerating 10nm rampGreat_Scott - Friday, April 24, 2020 - link
So they designed a new 10mn core but they can't produce many 10mn chips? How is this going to work?I call it now, <insert codename here> 14mm variant with the same 11th-generation tag will come out at the same time...
mode_13h - Friday, April 24, 2020 - link
Uh, Comet Lake is 14 nm and it's to be followed by Rocket Lake. So, that could certainly ship for mobile, as Comet Lake did.So, you're not wrong, but also not really going out on a limb, either. Certainly *could* happen.
yeeeeman - Friday, April 24, 2020 - link
Mobile will move entirely to 10nm. There will also be a 45w h series 8 core tigerlake cpu at some point.Deicidium369 - Saturday, April 25, 2020 - link
That would be fore the NUC11 most likely or NUC12Deicidium369 - Saturday, April 25, 2020 - link
Rocket Lake on laptops is called Tiger Lake.Deicidium369 - Saturday, April 25, 2020 - link
Uh 10nm has been shipping in volume since Last May/June. the 10nm+ process is well ramped up and Tiger Lake / Ice Lake Xeons, Agilex FPGAs. etc - quite a lot of 10nm shipping and plenty of 10nm+ capacity.You can call it whatever you want - Tiger Lake is Gen 11. Rocket Lake S - even though on 14nm - it is Tiger Lake under the hood - also Gen 11.
Gen has nothing to do with lithography.
Spunjji - Monday, April 27, 2020 - link
"In volume" is one of those terms that seems to be increasingly open to interpretation. It's certainly not at the volumes traditionally expected of Intel.drothgery - Tuesday, April 28, 2020 - link
It hasn't shipped in high enough volumes to let Intel give up on 14nm respins even just in mobile. But there are still a pretty huge number of them out there, or you wouldn't be able to get Ice Lake chips in sub-$1K laptops now (rather than just expensive premium models like my XPS 2-in-1) and it wouldn't be in fairly high-volume laptops like Apple's newest.Deicidium369 - Monday, May 4, 2020 - link
Well then by comparison AMD is nothing more than a paper launch since Ryzen 1.Korguz - Monday, May 4, 2020 - link
still better then intels 10nm availabilityDeicidium369 - Saturday, May 23, 2020 - link
than, not then. Intel's - the apostrophe (') indicated posession/ownership.Still, (commas are your friend) better than Intel's 10nm availability. (a period denotes the end of a sentence).
Korguz still not adding anything at all to the discussion at hand. Don't worry, keep trying. I always root for someone with sub par intelligence when they try to add something to the conversation. Keep trying, little buddy! We are all rooting for you. You CAN Do It!
Deicidium369 - Saturday, April 25, 2020 - link
10nm has been shipping for almost a year now - you need to find a new meme other than 10nm LOLtrenzterra - Friday, April 24, 2020 - link
Hopefully it won't launch together with a Meteor Lake or something on a 14+++++ process.yeeeeman - Friday, April 24, 2020 - link
Meteor lake is slated for 2022 and 7nm. You wanted to say rocket lake which is a desktop chip based on the same core design as tigerlake but built on 14nm.Deicidium369 - Saturday, April 25, 2020 - link
Know what's funny? TSMC had 16nm - made very tiny improvments and called it 14 and more small improvements and they call it 12.So instead of + - TSMC makes up new names to show to their customers - since Intel only makes and sells Intel - the 14+ being 12nm and the ++ being 11nm wasn't needed. If you think that the latest iteration of Intel 14nm bears any resemblance to it's first iteration - you are mistaken.
You kiddies need to find a new meme - 10nm LOL is DEAD, and the 14++++++++++ was dead quite a while back. What else ya got?
Spunjji - Monday, April 27, 2020 - link
This *isn't strictly correct*, though. Yes, TSMC rename processes that don't have substantial shrinks. Some of the changes Intel made to 14nm over time actually decrease density, though, and at least the first + was less of a physical process optimisation and more an optimisation of their ability to bin the resulting chips. You can fairly trivially undervolt an original quad-core mobile Skylake to run at Kaby Lake Refresh power levels.You keep stating one or two faces and then spinning them into this bizarre pro-Intel narrative.
You need to get less personally upset about people making light of Intel's manufacturing woes. They're nearly over, but you're still *really* pressed about it.
Deicidium369 - Monday, May 4, 2020 - link
I am not personally upset or actually even care one bit about any of this. Well, yes Intel HAD issues with 10nm - not any more. Kinda the inverse of TSMC having loser node after loser node and FINALLY getting on or 2 correct (16/14/12 and their 10nm class product "7nm"I dont take any of this personally or care what you have to say. I wish you were a little bit more informed and could have a real discussion.
Averant - Friday, April 24, 2020 - link
What I'd like to know is what that device he's holding is. Fascinating PCB with the prototype chip soldered on, by the looks of it.mode_13h - Friday, April 24, 2020 - link
It's a motherboard for an ultra-book or equivalent laptop.mode_13h - Friday, April 24, 2020 - link
Here's a teardown of a 2018 Macbook Air:https://www.youtube.com/watch?v=HMVDIMmbeoA
Deicidium369 - Saturday, April 25, 2020 - link
Yeah not unlike the one for Dell 13 2-in-1s with Ice Lake. I dont think there will be much that needs to be changed to move to TGL.lmcd - Friday, April 24, 2020 - link
Do we know yet if Intel's Xe-LP graphics will continue to offer GPU virtualization to consumers? It's one of the secret upsides of an Intel iGPU and I'm concerned it will disappear as Intel gets more serious about its graphics (except for professional SKUs).mode_13h - Friday, April 24, 2020 - link
If you sift around through the oneAPI docs enough, you might find some answers. That's a pretty deep question, though.lmcd - Friday, April 24, 2020 - link
It's currently possible to use Broadwell or newer iGPUs in a virtualized setting with QEMU without *too* much of a fight, from what I've read. I'm planning on trying this configuration with a macOS virtual machine but haven't quite gotten to it.Niche need for sure but if it ends up being supported, I'll be thrilled.
mode_13h - Friday, April 24, 2020 - link
> with any luck, Intel’s 4+ years of playing with 10nm may finally pay some better dividends as they bring up their latest process.It's funny you should use that analogy. If Intel hadn't been so intent on paying out dividends (and doing buybacks), their 10 nm might've come on sooner and stronger.
slide> * Committed to maintaining the dividend
slide> * Repurchased $4.2B in shares
Well, it's good to see they've got their priorities straight. Uh...
yeeeeman - Friday, April 24, 2020 - link
Yeah, they stopped pumping money into r&d and this was a bad decision.Yojimbo - Friday, April 24, 2020 - link
You're being sarcastic?Yojimbo - Friday, April 24, 2020 - link
Intel's R&D by year:2015 - $12.1B
2016 - $12.7B
2017 - $13.1B
2018 - $13.5B
2019 - $13.4B
They did spend a lot in buyback the last couple of years but it has nothing to do with their R&D spend. Their issues were not a result of not throwing enough money around. It was poor execution, and to fix it takes better execution and time, not just throwing around money.
As far as the buybacks and dividends, they have had huge profits. Giving back to investors is what they should be doing. They are already large with many pots on the fire. There comes a point where diversifying into more businesses will only dilute the company and won't provide the best value for shareholders. Giving returns to shareholders isn't corruption, it's sharing the profits of a successful business.
yeeeeman - Saturday, April 25, 2020 - link
Yeah, I was sarcastic. I know it is a combination of many factors. I would add to what you said, bad management. They got used to great execution they even created the tick tock model. When 22nm ramped harder than previous nodes, they probably didn't adjust fab plans. 14nm was even worse but they still didn't learn their lesson because with 10nm they just went overboard with the features. any sane person would look at the past two nodes and say...let's learn something and do this one differently. Anyway...Spunjji - Monday, April 27, 2020 - link
100% with yeeeeman here. They let marketing set the roadmaps instead of engineering, and I get the impression there was an increasingly strong "yes culture" in middle-management that obfuscated the extent of the issues they were having in the foundry until it was too late, at which point they were deep into saving face at all costs.Yojimbo - Monday, April 27, 2020 - link
Could be, but I think we're in the realm of pure speculation here. For all we know it was the engineering team leaders saying "push, push, we can do this."Spunjji - Monday, April 27, 2020 - link
The narrative about management woes has been corroborated by several long serving ex-Intel engineers - so while the usual cautions about trusting anecdotal evidence from former employees come into play, it does sound (and look) a lot like there were serious issues at the middle-management level that were at least significant contributing factors to their current difficulties.name99 - Monday, April 27, 2020 - link
For anyone who wants to collaborate this compare:https://www.thelayoff.com/intel
https://www.thelayoff.com/apple
The Intel complaints are pretty much all about how their managers suck.
The Apple complaints (ever since Apple picked up the modem division) are ALSO pretty much all about how the Intel managers suck!
Spunjji - Tuesday, April 28, 2020 - link
lol. Says it all, really.I did get a kick out of Apple buying their junk modem division. Apple will be taking revenge for that one soon enough when they kick Intel out of their low-end laptops.
Deicidium369 - Saturday, April 25, 2020 - link
All comes down to trying to make Gen 1 of 10nm 2.7X density increase instead of 2x. 10nm+ reaches that 2.7x density target. I also think that Cobalt was a lot more challenging - something that neither Samsung or TSMC has started on. So this time was used to benefit long term. Besides they could not make enough 14nm to fill all the orders, so 10nm would have been cool several years ago, but from a business point of view - un needed.Depends on the other business units being expanded - 5G was sold to Apple, their Networking unit is very strong (even selling on most AMD systems) as are the Optane unit. So dilution is not much of an issue - this all plays into the taking control of more of the TAM for silicon.
Spunjji - Monday, April 27, 2020 - link
"Besides they could not make enough 14nm to fill all the orders, so 10nm would have been cool several years ago, but from a business point of view - un needed."How would having additional manufacturing capacity from a newer node - and being able to continue out-competing their closest rival on both transistor density and CPU core design - ever have been "un-needed"? 🤨
Deicidium369 - Monday, May 4, 2020 - link
I know in the AMD world lithography is some magical panacea - but it's not.You miss the point like a short outfielder misses a pop fly. Business only cares about revenues - and of your older much more optimized platform is selling in volumes you cannot keep up with - adding another newer products on the new process - would make little sense.]
From an enthusiast point of view - I would say that it WAS needed...
mode_13h - Saturday, April 25, 2020 - link
Intel's pretty notorious for making repeated and painful job cuts.I don't have the data to back this up, but I think you should try to compare that R&D spend to a basket of companies that cover the equivalent business activities and markets to where Intel plays. I don't think your picture would look so rosy, then.
My sense is that each new manufacturing node is significantly more expensive than the last. And here's where Intel probably fell down. Why don't you plot TSMC's R&D spend, over the same time? If you can get Samsung's semiconductor manufacturing R&D, put that up too.
I never said that dividends were corruption, but I think you really can't argue that it's actually *good* for operations to do it, regardless of whether a company is executing to plan. When a company is having problems with execution, maybe the root cause isn't money, but often some extra spend can still help them get through a rough patch.
Oxford Guy - Sunday, April 26, 2020 - link
"My sense is that each new manufacturing node is significantly more expensive than the last."A TSMC person, a few years ago, stated how many design rules needed to be used for the different nodes. The number of design rules (complexity of design) increased very rapidly with node shrinkage.
Deicidium369 - Saturday, April 25, 2020 - link
Dividends are not vital but are expected - they do not horde cash overseas like ApplePriorities are to the invenstors / shareholders
Pretty sure paying dividends are not the reason that Ice Lake was released when it was - was released when it was ready
Oxford Guy - Saturday, April 25, 2020 - link
It's about time for Anandtech to post a detailed article about the actual node sizes involved in all of the different CPUs on the market and coming soon.The fake nm naming stuff is intentionally confusing for consumers. It's not good enough to just relegate the truth to some random comments in articles.
Wilco1 - Saturday, April 25, 2020 - link
What really matters is the actual transistor density of real devices. As opposed to marketing.TSMC 7nm: 90.9 million transistors / mm^2 (Kirin 990 5G)
Intel 10nm: 49.4 MT/mm^2 (Lakefield compute die)
So that shows for mobile SoCs, Intel 10nm is nowhere near TSMC in real density.
https://en.wikipedia.org/wiki/Transistor_count
https://fuse.wikichip.org/news/3417/a-look-at-inte...
yeeeeman - Saturday, April 25, 2020 - link
Efficiency for the Intel node is also much worse than tsmc 7nm.yeeeeman - Saturday, April 25, 2020 - link
Basically the revised 10nm node is just a bit denser than 14nm. Much much different than what they intended in the first place where they stated 100mtr/ mm2, better than tsmc 7nm.Deicidium369 - Saturday, April 25, 2020 - link
in what way? power usage? Power usage is quite efficient on Ice Lake and assuming Tiger Lake.and, no 10nm+ is not a bit denser than 14nm - 10+ reaches the initial target of 2.7x density increase.
Intel traditionally has multiple architectures - some frequency optimized (current 5GHz+) or power optimized (lakefield, Ice Lake) or density optimized. The 10nm is focused on power efficiency, 10nm+ will focus on density and (Ice Lake Servers) and frequency is likely never to reach 14nm levels - but then again neither can any node from TSMC
AMD's Zen 2 compute complex is 30-36M/mm^2 Intel 14nm is 37-45M/mm^2. Source - https://www.reddit.com/r/Amd/comments/ecca4s/trans...
Spunjji - Monday, April 27, 2020 - link
Your AMD numbers are wrong. Zen 2's *compute complex* is 52.7MTr/mm^2. You quoted the density of the *entire chip*, which includes the 14nm IO die. Renoir - which is entirely 7nm - is up at 63.33MTr/mm^2.https://www.reddit.com/r/Amd/comments/fraqll/amd_s...
As for your claims about Intel's 10nm+ density, I literally can't find any solid information that would confirm or refute it. You're effectively quoting their original 10nm marketing slides verbatim and expecting that to hold up in reality.
name99 - Monday, April 27, 2020 - link
Intel's best case real world scenario 10nm density numbers are for Lakefield's compute die. This die omits all those large IO transistors, and has no requirements to be blazing fast. Even so, yeah, half the density of TSMC...You can find the raw numbers here if you want to do your own calculation:
https://fuse.wikichip.org/news/3417/a-look-at-inte...
Spunjji - Tuesday, April 28, 2020 - link
Cheers for the sauce. I'm amazed they missed their advertised numbers by that far, especially given the relatively small caches and offloaded IO. Given the purported numbers for Ice Lake, though, I do wonder if there's something missing from the story here.Foveros looks quite promising as a future technology. Right now, though, it's amusing they had to do all of that just to reach comparable board space requirements (but not power/performance parity) with Apple's ARM designs.
Deicidium369 - Monday, May 4, 2020 - link
That is not my Reddit post. I don't post on Reddit and don't even have an acct thereDeicidium369 - Saturday, April 25, 2020 - link
what is that like the 4th or 5th iteration of TSMC "7nm"?Lakefield is Intel 10nm which is only a 2x increase on 14nm - 10nm+ (Tiger Lake) is 2.7x density.
That is TSMCs high density line - the Zen 2 CCX is less than 40M/mm^2.
see source linked below
Wilco1 - Saturday, April 25, 2020 - link
Kirin 990 5G uses the 2nd generation 7nm process, N7+ (EUV). Lakefield uses 10nm+ as Intel's slide clearly shows.Zen 2 chiplet is 52.7 MT/mm^2 using TSMC's first-gen 7nm. 14nm Xeons get just 15.8 MT/mm^2 (so about 3.3 times less dense).
We'll see how Tiger Lake does, but given it must use far more large and IO transistors than Lakefield, it'll be lucky to achieve 40 MT/mm^2, or about 2.5x scaling. Zen 3 will be out at the same time as Tiger Lake, if it uses 7nm+, density should improve to 60 MT/mm^2.
jeremyshaw - Monday, April 27, 2020 - link
To be entirely fair, the Zen2 chiplet has very little external I/O and a LOT of SRAM cache (basically best case for density), while the 14nm Xeons (assuming LGA2066 or LGA3xxx, but even the client LGA11xx/1200 somewhat applies) have a lot of I/O, much of it that has to be driven over longer distances (which drives up feature size considerably). Intel chips also come with less cache vs their AMD counterparts. So 3.3x less dense isn't exactly a straight comparison.Of course, AMD's design is made with large caches in mind, so it's not some sort of unfair comparison overall, since AMD's design uses that large cache to accommodate their slight latency penalty for having off-die I/O.
As for Tiger Lake, it has to accommodate a GPU, which are even more notoriously cache-light, lowering the density metric even further. Zen2 chiplet does not. The more interesting comparison, IMO, is Ice Lake & Tiger Lake density vs Zen2+Vega & Zen2/3+RDNA2. In these cases, AMD lowers their total cache by almost half, integrates a GPU, and also has to drive external I/O (DRAM, PCIe, USB, SATA, DP, eDP, etc). The comparison would have been less fair in the past, when the Intel U series chips had only DRAM, DP/eDP, and 4x PCIe lanes to account for (everything else was from the external chipset), but those are now joined by 4 TB3 controllers (each with 4 high speed differential pairs) basically not too dissimilar from having the SerDes for another 8 PCIe lanes, operating at a link rate between PCIe 4 and 5. Ice Lake also increases per core cache, too.
I think it's a fair comparison, as a result. Zen2 Renoir (APU) increased density to about 63.3MT/mm^2 (up from 52.7 MT/mm^2 on the Zen2 chiplet), this is inspite of decreased cache and increased I/O. Intel is treating Ice Lake-U transistor density like a trade secret, but they claim 7B transistors and measurements from Anandtech show 122.5mm^2 CPU size, giving ~57.1 MT/mm^2. Still not a straight comparison, as AMD doesn't seem to have made a low end, 7nm quad core APU this gen, and Intel only makes a quad core flagship 10nm product. Intel also still relies on a southbridge to drive the SATA, PCIe hub, USB, etc. It is within the ballpark for Intel's 10nm "high performance" variant's density claim (somewhere in the 60+ MT/mm^2 range, iirc).
Of course, this all has another problem: similar to "10nm" and "7nm" marketing names not meaning a whole lot, everyone measures "transistor count" differently, and even for the same company, the method has changed over time.
In the end, it basically means the density measurements we can get/make are almost meaningless.
Spunjji - Monday, April 27, 2020 - link
Just wanted to congratulate you on the quality of this post. Cutting through some nonsense here 👍Spunjji - Monday, April 27, 2020 - link
I'm not sure the comparison between Kirin and Lakefield is necessarily that enlightening, given how much sustained effort ARM put into optimising their designs for area efficiency on a given process technology.Renoir vs. Ice Lake ought to be a better point of comparison. Renoir runs 63.3MTr / mm^2 - I couldn't actually find numbers for ICL, though; just repetitions of Intel's original 100MTr / mm^2 claim which I believe was for NAND cells. I know it's lower than that for ICL on 10nm+, but have no idea by how much.
Wilco1 - Tuesday, April 28, 2020 - link
Lakefield is aiming to be a small low power chip like Kirin. It isn't possible to get a perfect comparison, but comparing mobile vs mobile and server vs server is quite reasonable. Either way, Renoir vs Lakefield doesn't look too good either. Despite Renoir being high-end and having more IO, it still has much better density.Bottom line - the few 10nm numbers we have show it is well behind first-gen 7nm despite all the marketing claims of 100MT/mm^2.
Spunjji - Tuesday, April 28, 2020 - link
That's a fair point - I just figure that what's effectively a first-gen product on new manufacturing tech (Foveros) is probably still not anywhere close to optimal. But you're right, that's the class of product it's effectively competing against...Definitely agreed on the last point. 100MT/mm^2 was always an aspiration at best, and anyone repeating it seriously now needs their ears rinsed out.
Everett F Sargent - Monday, May 4, 2020 - link
So, it has now been FIVE YEARS since Intel released 14mn desktop silicon (Q2'15 65W and 95W) ...https://ark.intel.com/content/www/us/en/ark/produc...
So when will Intel finally release desktop (or better, meaning at least 65W-95W parts) at any process node below 14nm? I am guessing Q2'22 or SEVEN YEARS after their 1st 14mn desktop parts.
watzupken - Monday, May 4, 2020 - link
If the first picture is the actual die size, I think it looks quite a lot bigger than the current Ice Lake U. Having used an Ice Lake U laptop, I feel that the 10nm seems to be closer to 14nm than 7nm. In fact, it seems to have regressed due to the amount of refinement over the years. Taking a quick comparison between Ice Lake and Whiskey Lake, the base clock have dropped significantly, and in some case, increased power requirement. In addition, the transition from 14nm to 10nm don't seem to help with the thermals. In my IdeaPad 5, I observed temps climbing up to 94 degs. While temp is a subjective metric highly dependent on the manufacturer's cooling solution and configuration, still 94 degs is the highest I have seen in a laptop. Even my older Ryzen 2500U don't go above 84 degs under sustained load.