Shouldn't memory for Kaby Lake systems be able to run on Ryzen systems too? Both chipsets/cpu's support dual channel DDR4, same number of channels, max memory, etc.
AMD should obviously not be at the same level of XMP support as Intel. It's like expecting OpenOffice to be able to render .docx Word documents as well as Microsoft Office's Word application, especially when both .docx files and XMP aren't necessarily open standards.
Yes, AMD has some issues with supporting DRAM at speeds over 3000MHz, but in reality, the bulk of DDR4 memory sales are (ECC or Non-ECC) memory ranging from speeds between 2133 -> 3000MHZ anyways, and they capture the majority of users in that range.
Would it be nice to be able to run DDR4 4000MHz+ memory on an AMD platform? Sure. But likewise I'd also enjoy being able to use ECC memory on a Z270 platform, which isn't something Intel supports.
If I'm going to make an expensive memory purchase, this is my personal preference on order of importance:
ECC > Capacity > Speed
For example, if I can budget $200 for memory on a specific PC build, I'd rather go for ECC memory first, then in the highest capacity I can find, then at the best speed I can find.
AMD's memory platform works just fine for me, but I know other users have different preferences for memory and would rather splurge big bucks for lower capacity, non-ECC, RAM rated for high speeds.
It all sounds like excuses to me. XMP or no XMP there's no excuse not to handle the higher clock speeds particularly with a newer product on the market.
Wanna let me know Intel's "excuse" for not supporting ECC on their mainstream product line for over a decade? Because the technology's there, they're choosing not using it.
Meanwhile, AMD's basically trying reverse engineer how Intel's XMP works and trying to get to the same level of DDR4 support from where before they had been stuck at DDR3 RAM in their previous chipsets.
You might think I'm being an AMD apologist here for explaining how the XMP situation is in reality, but you really need to take a look at team blue (intel) here and look at what apologist still thinks it's OK that ECC memory isn't supported across all their platforms in the year 2017.
AMD could do ECC memory support on mainstream platforms, why can't Intel?
ECC memory costs more with no added benefit for most users... Most users care about capacity then speed, even if ECC was an option it would come last.
XMP is just a table with some more parameters than SPD, nothing extraordinary... AMD wants to use it cause it's already here and AMD doesn't have enough market share to make memory makers adopt it.
Likewise, once you can fully saturate the program's needs for memory, capacity doesn't matter, either. (16GB is plenty for just about any application today, and you could argue that 8GB is plenty, too, but has a bit more issues when dealing with concurrency with lots of applications open at a time, while 32GB or more is overkill unless you _KNOW_ you're running a memory-intensive application.)
Meanwhile ECC memory, when utilized, gives a system rock-solid stability against memory-based errors. This can prevent a random application crash that happened out of the blue, or a random system hangup, internet browser that suddenly became unresponsive, blue screen, etc.
Those who haven't used ECC don't understand its importance over speed and capacity and somehow continue to perpetuate this meme that DRAM speed offers tangible benefits to gamer FPS/$ ratios, when you should just be spending that extra money you're wasting on faster DRAM on a faster processor or GPU instead.
"It's been said again and again and again that faster DRAM speed has negligible results for PCs." That's why I put memory capacity first... negligible is better than NOTHING, also for gamers it MAKES a difference albeit small.
I watch Linus on Youtube too... that test compared RAM speed on a GPU bond environment, I could bet that if he switched the I7 for an I3 we would still see no difference on games at least. http://www.techspot.com/article/1171-ddr4-4000-mhz... http://www.eurogamer.net/articles/digitalfoundry-2... Anyways the benefit from faster RAM depends on the system and game... but it clearly exists. Also what about people that are using iGPU and hence using the main memory for graphics too?...
When I say capacity I don't mean have as much as you can... but have at least the minimum you NEED.
My PCs have a rock solid stability without ECC. Most sources of instability on a system are due to software/hardware problems and there's nothing ECC can do about it. ECC only protects the RAM against bit flip due to external forces (cosmic rays and what not) and internal RAM problems. Your Pugget link compares the quality of ECC against non-ECC Ram, ECC is designed to be used on critical systems so I guess RAM producers will go an extra mile checking the Ram sticks, it's like comparing budget RAM against premium (this assuming premium is not just the price). They don't really show any data on how ECC makes a difference. From your source "So while ECC RAM is certainly important for servers and systems with high-value data, non-ECC RAM is more than stable enough for use in most home or work systems."
People don't care about ECC cause people don't really get the problems ECC is supposed to address.
>People don't care about ECC cause people don't really get the problems ECC is supposed to address.
No, people don't care about ECC because they're under the assumption that they never see cosmic ray events and that all faults seen when using their computer are just software based. They then mistakenly assume that MOAR speed and MOAR capacity beyond what is necessary offers tangible benefits when, depending on the workload, you hit a plateau where more speed or capacity is not inherently useful to your workload. 24/7 stability provided by ECC can be useful to anyone, even grandma and grandpa who only use Internet Explorer to send/receive pictures of grandkids to friends and family, where a cosmic ray event induced memory error can cause system instability.
Cosmic ray events aren't as rare as you might think, and even though you may not _see_ cosmic ray events, doesn't mean they don't have an effect on your system.
"Studies by IBM in the 1990s suggest that computers typically experience about one cosmic-ray-induced error per 256 megabytes of RAM per month", attributed to this source: https://www.scientificamerican.com/article/solar-s...
A PC equipped with 8GB of memory running 24/7 for a month will incurr, on average, about 32 cosmic ray events that result in a bit-flip in a DRAM module's register. If you run the PC 12 hours a day, then ~16 events will happen, or if you run a 16GB memory PC for 12 hours a day ~32 events will happen.
If you're noticing the correlation, the higher the memory capacity that the PC has and the more often it's running, the more cosmic ray events will incur a bit-flip, and therefore affect the system's performance and stability over a given measure of time, such as a month.
Likewise, you may think, "But hey! That's a 20 year old statistic, that can't be right today with how much denser and more electronically stable DRAM modules and CPUs of modern day." Well no, because according to this article, Intel states the problem increases in prevalence the smaller the integrated circuit becomes, so the 20 year old statistic is actually more of a best-case scenario: https://www.newscientist.com/blog/technology/2008/...
So yes, it very much is possible that the time your browser completely locks up randomly even though you weren't doing anything is due to a cosmic ray event flipping a bit in memory where your browser was cached and when the processor proceeds to fetch this information, the process locks up and looks "stuck" to the user. It might look like a software error, but if the exact same scenario isn't reproducible exactly, then chances are that YES it was a memory error and that YES it's possible this memory error was caused by a cosmic ray event and that YES ECC (registered) memory may have prevented this issue from ever occurring.
And then we go back to the initial topic at hand; Yes, ECC support on the platform gives users the CHOICE to purchase ECC memory at a ~25% premium over non-ECC memory of the same capacity and speed. AMD offers that CHOICE to users. Intel does NOT.
What's your Intel apologist excuse for Intel not providing that CHOICE for mainstream platform users for over a decade?
Not that Windows 10 crashes all that often anyway, but there is still a noticeable difference between workstation-grade hardware - which includes ECC memory - and a regular consumer-grade platform.
Not that the consumer grade hardware crashes often these days, but workstation/server grade crashes noticeably less. As in, pretty much never. I don't think my home workstation has ever gotten a BSOD.
I don't know how much of that is specifically attributable to the ECC memory, but it does seem like memory errors are responsible for a lot of crashes when they do occur.
ryzen uses a strange block multiplier to overclock memory to xmp speeds. this in change means it is more difficult by default to get memory working on ryzen. they even have a different type of south bridge interface! yikes. it will take about a year or so to get 3600mhz to run normally with micro code. just moar waitin.
They should put some extra transistors at connecting modules on L3 cache level rather than on MC level. The MC is very slow compared to L3 cache, they did that because it was easier as both modules are connected to the MC, so it can additionally act as an inter module communication hub. But it is very slow, which has a pronounced effect on stuff like games. Number crunching is not really affected as core affinities do not fluctuate so much plus saturating the pipeline helps to mask out inter module access latency.
It isn't slow; there's a higher average latency penalty for crossing CCXs, but if you use the wider bus width properly, you can fit more data with less penalty if you need cache coherency between CCX modules for dependent datasets. The NB does run at memory speed, but is 256-bit wide, which is how it provides good bandwidth and excellent power savings.
If you want actual numbers, 40ns to communicate to cores within a CCX through L3, 120ns to communicate between cores in a different CCX, which is an 80ns average.
That number isn't really that bad. As the memory's clock speed rises, the amount of time each cycle takes to complete decreases. More cycles may not actually mean higher latency. Though I do agree with the general idea of aggressively lowering latency in addition to raising memory speed, there are physical limits that are difficult to overcome.
It's actually faster than DDR4-3200 14-14-14. There the latency is 1/3.2 * 14*2 = 8.75 ns, whereas for DDR4-4333 19-19-19 it's 1/4.333 * 19*2 = 7.03 ns.
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.
23 Comments
Back to Article
HomeworldFound - Saturday, April 15, 2017 - link
I'd take them in quad channel and in titanium white.Chriz - Saturday, April 15, 2017 - link
Shouldn't memory for Kaby Lake systems be able to run on Ryzen systems too? Both chipsets/cpu's support dual channel DDR4, same number of channels, max memory, etc.ZeDestructor - Saturday, April 15, 2017 - link
Apparently XMP support is iffy on AMD's end of things...JoeyJoJo123 - Monday, April 17, 2017 - link
XMP is Intel developed; https://en.wikipedia.org/wiki/Serial_presence_dete...AMD should obviously not be at the same level of XMP support as Intel. It's like expecting OpenOffice to be able to render .docx Word documents as well as Microsoft Office's Word application, especially when both .docx files and XMP aren't necessarily open standards.
Yes, AMD has some issues with supporting DRAM at speeds over 3000MHz, but in reality, the bulk of DDR4 memory sales are (ECC or Non-ECC) memory ranging from speeds between 2133 -> 3000MHZ anyways, and they capture the majority of users in that range.
Would it be nice to be able to run DDR4 4000MHz+ memory on an AMD platform? Sure. But likewise I'd also enjoy being able to use ECC memory on a Z270 platform, which isn't something Intel supports.
If I'm going to make an expensive memory purchase, this is my personal preference on order of importance:
ECC > Capacity > Speed
For example, if I can budget $200 for memory on a specific PC build, I'd rather go for ECC memory first, then in the highest capacity I can find, then at the best speed I can find.
AMD's memory platform works just fine for me, but I know other users have different preferences for memory and would rather splurge big bucks for lower capacity, non-ECC, RAM rated for high speeds.
HomeworldFound - Tuesday, April 18, 2017 - link
It all sounds like excuses to me. XMP or no XMP there's no excuse not to handle the higher clock speeds particularly with a newer product on the market.JoeyJoJo123 - Tuesday, April 18, 2017 - link
Wanna let me know Intel's "excuse" for not supporting ECC on their mainstream product line for over a decade? Because the technology's there, they're choosing not using it.Meanwhile, AMD's basically trying reverse engineer how Intel's XMP works and trying to get to the same level of DDR4 support from where before they had been stuck at DDR3 RAM in their previous chipsets.
You might think I'm being an AMD apologist here for explaining how the XMP situation is in reality, but you really need to take a look at team blue (intel) here and look at what apologist still thinks it's OK that ECC memory isn't supported across all their platforms in the year 2017.
AMD could do ECC memory support on mainstream platforms, why can't Intel?
Strunf - Tuesday, April 18, 2017 - link
ECC memory costs more with no added benefit for most users...Most users care about capacity then speed, even if ECC was an option it would come last.
XMP is just a table with some more parameters than SPD, nothing extraordinary... AMD wants to use it cause it's already here and AMD doesn't have enough market share to make memory makers adopt it.
JoeyJoJo123 - Tuesday, April 18, 2017 - link
>ECC memory costs more with no added benefit for most users...Look, two can play that game.
>Faster memory costs more with no added benefit for most users...
https://www.youtube.com/watch?v=dWgzA2C61z4
It's been said again and again and again that faster DRAM speed has negligible results for PCs.
https://www.youtube.com/watch?v=dWgzA2C61z4
Likewise, once you can fully saturate the program's needs for memory, capacity doesn't matter, either. (16GB is plenty for just about any application today, and you could argue that 8GB is plenty, too, but has a bit more issues when dealing with concurrency with lots of applications open at a time, while 32GB or more is overkill unless you _KNOW_ you're running a memory-intensive application.)
https://www.youtube.com/watch?v=dWgzA2C61z4
Meanwhile ECC memory, when utilized, gives a system rock-solid stability against memory-based errors. This can prevent a random application crash that happened out of the blue, or a random system hangup, internet browser that suddenly became unresponsive, blue screen, etc.
https://www.pugetsystems.com/labs/articles/Advanta...
Educate yourself.
Those who haven't used ECC don't understand its importance over speed and capacity and somehow continue to perpetuate this meme that DRAM speed offers tangible benefits to gamer FPS/$ ratios, when you should just be spending that extra money you're wasting on faster DRAM on a faster processor or GPU instead.
fanofanand - Tuesday, April 18, 2017 - link
*mic drop*Strunf - Wednesday, April 19, 2017 - link
"It's been said again and again and again that faster DRAM speed has negligible results for PCs."That's why I put memory capacity first... negligible is better than NOTHING, also for gamers it MAKES a difference albeit small.
I watch Linus on Youtube too... that test compared RAM speed on a GPU bond environment, I could bet that if he switched the I7 for an I3 we would still see no difference on games at least.
http://www.techspot.com/article/1171-ddr4-4000-mhz...
http://www.eurogamer.net/articles/digitalfoundry-2...
Anyways the benefit from faster RAM depends on the system and game... but it clearly exists. Also what about people that are using iGPU and hence using the main memory for graphics too?...
When I say capacity I don't mean have as much as you can... but have at least the minimum you NEED.
My PCs have a rock solid stability without ECC. Most sources of instability on a system are due to software/hardware problems and there's nothing ECC can do about it. ECC only protects the RAM against bit flip due to external forces (cosmic rays and what not) and internal RAM problems.
Your Pugget link compares the quality of ECC against non-ECC Ram, ECC is designed to be used on critical systems so I guess RAM producers will go an extra mile checking the Ram sticks, it's like comparing budget RAM against premium (this assuming premium is not just the price). They don't really show any data on how ECC makes a difference.
From your source "So while ECC RAM is certainly important for servers and systems with high-value data, non-ECC RAM is more than stable enough for use in most home or work systems."
People don't care about ECC cause people don't really get the problems ECC is supposed to address.
BrokenCrayons - Wednesday, April 19, 2017 - link
Excellently put Strunf!JoeyJoJo123 - Wednesday, April 19, 2017 - link
>People don't care about ECC cause people don't really get the problems ECC is supposed to address.No, people don't care about ECC because they're under the assumption that they never see cosmic ray events and that all faults seen when using their computer are just software based. They then mistakenly assume that MOAR speed and MOAR capacity beyond what is necessary offers tangible benefits when, depending on the workload, you hit a plateau where more speed or capacity is not inherently useful to your workload. 24/7 stability provided by ECC can be useful to anyone, even grandma and grandpa who only use Internet Explorer to send/receive pictures of grandkids to friends and family, where a cosmic ray event induced memory error can cause system instability.
Cosmic ray events aren't as rare as you might think, and even though you may not _see_ cosmic ray events, doesn't mean they don't have an effect on your system.
https://en.wikipedia.org/wiki/Cosmic_ray#Effect_on...
"Studies by IBM in the 1990s suggest that computers typically experience about one cosmic-ray-induced error per 256 megabytes of RAM per month", attributed to this source: https://www.scientificamerican.com/article/solar-s...
A PC equipped with 8GB of memory running 24/7 for a month will incurr, on average, about 32 cosmic ray events that result in a bit-flip in a DRAM module's register. If you run the PC 12 hours a day, then ~16 events will happen, or if you run a 16GB memory PC for 12 hours a day ~32 events will happen.
If you're noticing the correlation, the higher the memory capacity that the PC has and the more often it's running, the more cosmic ray events will incur a bit-flip, and therefore affect the system's performance and stability over a given measure of time, such as a month.
Likewise, you may think, "But hey! That's a 20 year old statistic, that can't be right today with how much denser and more electronically stable DRAM modules and CPUs of modern day." Well no, because according to this article, Intel states the problem increases in prevalence the smaller the integrated circuit becomes, so the 20 year old statistic is actually more of a best-case scenario: https://www.newscientist.com/blog/technology/2008/...
So yes, it very much is possible that the time your browser completely locks up randomly even though you weren't doing anything is due to a cosmic ray event flipping a bit in memory where your browser was cached and when the processor proceeds to fetch this information, the process locks up and looks "stuck" to the user. It might look like a software error, but if the exact same scenario isn't reproducible exactly, then chances are that YES it was a memory error and that YES it's possible this memory error was caused by a cosmic ray event and that YES ECC (registered) memory may have prevented this issue from ever occurring.
And then we go back to the initial topic at hand; Yes, ECC support on the platform gives users the CHOICE to purchase ECC memory at a ~25% premium over non-ECC memory of the same capacity and speed. AMD offers that CHOICE to users. Intel does NOT.
What's your Intel apologist excuse for Intel not providing that CHOICE for mainstream platform users for over a decade?
twtech - Monday, April 24, 2017 - link
Not that Windows 10 crashes all that often anyway, but there is still a noticeable difference between workstation-grade hardware - which includes ECC memory - and a regular consumer-grade platform.Not that the consumer grade hardware crashes often these days, but workstation/server grade crashes noticeably less. As in, pretty much never. I don't think my home workstation has ever gotten a BSOD.
I don't know how much of that is specifically attributable to the ECC memory, but it does seem like memory errors are responsible for a lot of crashes when they do occur.
austinsguitar - Saturday, April 15, 2017 - link
ryzen uses a strange block multiplier to overclock memory to xmp speeds. this in change means it is more difficult by default to get memory working on ryzen. they even have a different type of south bridge interface! yikes. it will take about a year or so to get 3600mhz to run normally with micro code. just moar waitin.Lolimaster - Saturday, April 15, 2017 - link
They should just release fast memory only for Ryzen, the only cpu that actually makes use of the faster speed because of their CCX design.ddriver - Sunday, April 16, 2017 - link
They should put some extra transistors at connecting modules on L3 cache level rather than on MC level. The MC is very slow compared to L3 cache, they did that because it was easier as both modules are connected to the MC, so it can additionally act as an inter module communication hub. But it is very slow, which has a pronounced effect on stuff like games. Number crunching is not really affected as core affinities do not fluctuate so much plus saturating the pipeline helps to mask out inter module access latency.Lolimaster - Sunday, April 16, 2017 - link
Ryzen still performs better in games using Win7 ("unsupported") vs Win10.Aside from not having DX12 the effect on games is greatly minimized.
JasonMZW20 - Sunday, April 16, 2017 - link
It isn't slow; there's a higher average latency penalty for crossing CCXs, but if you use the wider bus width properly, you can fit more data with less penalty if you need cache coherency between CCX modules for dependent datasets. The NB does run at memory speed, but is 256-bit wide, which is how it provides good bandwidth and excellent power savings.If you want actual numbers, 40ns to communicate to cores within a CCX through L3, 120ns to communicate between cores in a different CCX, which is an 80ns average.
MrSpadge - Wednesday, April 19, 2017 - link
80 additional ns.. that would be longer than an Intel IMC needs to access main memory, which should be far slower than inter-chip communication!NeatOman - Sunday, April 16, 2017 - link
That's a high CL right there :-/Bragging rights at that point, I'll wait for a sensible kit with much tighter timings.. but ima want me some 4GHz ram doe.
BrokenCrayons - Monday, April 17, 2017 - link
That number isn't really that bad. As the memory's clock speed rises, the amount of time each cycle takes to complete decreases. More cycles may not actually mean higher latency. Though I do agree with the general idea of aggressively lowering latency in addition to raising memory speed, there are physical limits that are difficult to overcome.MrSpadge - Wednesday, April 19, 2017 - link
It's actually faster than DDR4-3200 14-14-14. There the latency is 1/3.2 * 14*2 = 8.75 ns, whereas for DDR4-4333 19-19-19 it's 1/4.333 * 19*2 = 7.03 ns.andychow - Tuesday, April 25, 2017 - link
The JEDEC standard is pretty much out of the window.