The DC D series drives aren't going to be using 3D NAND initially, so there's no particular reason to expect a big capacity boost. The DC P3320 will eventually grow beyond 2TB, but those details haven't been announced.
As far as I can tell Samsung's only DC product at 2.5" NVMe PCIe is the PM953. It goes up to 1.92TB and the specs are 1GB/s read, 870MB/s write for the largest capacity (google MZQLV1T9HCJM). Only the PCIe cards from Samsung go up to 6TB (6GB/s read, 2GB/s write). Everything else they offer is SAS or SATA.
Or were you comparing Intel's DC SSDs to Samsung's consumer SSDs? In this case it's because you can sell anything to a consumer as long as you put a big number on the box, that you can reach only in some very specific laboratory conditions and only after you sacrifice a virgin for the speed gods. Reliability is another major difference between the two.
Otherwise you'd be pressed to ask yourself why do those guys with the datacenter constantly go for the products that have worse specs than what you can order from Amazon but many times the price...
Keep in mind Intel's rating here is at less than 10% variation IOP to IOP; beating the drive at 100% throughput continuously. Samsung's SM961 is rated at peak IOPs over a very short period, at least if it behaves anything like the 950 Pro. (see http://www.anandtech.com/show/9702/samsung-950-pro... and look at the IOPS graphs; and how the steady-state IOPS of Intel's 750 is around 4x what the 950 Pro manages)
As a consumer drive the Samsung is probably better. As a SQL Server drive, the Intel drive is probably better. Depends on your use case.
SQL and any server running a database (even a mail server) will have tons of tiny random writes, even in log files. The problem is since most drives don't store any user data in dram cache, and only TLC drives have an SLC "buffer" all of this random IO will wear a non-enterprise drive that lacks firmware to expect this type of environment. Enterprise drives often sacrifice a little speed by delaying buffer flushes to NAND or actually caching tiny IO data (512kb before a page write) to not hammer the NAND so much. Even Samsung has the PM841DC (or something like that) which is identical to the 840 Pro with vastly different firmware. It's also slower by like 100MB/sec!
Drives that cache user data to dram cache, with the exception of many Intel drives that supposedly only use it to cache the indirection table, usually have power loss protection. Even Anandtechs review of the Intel drives suspects a little user data gets cached to DRAM though.
The SM961 is an OEM drive for the client market, so it's not comparable. Client specifications are often higher because client IO workloads are bursty, making it okay to quote maximum performance, not sustained.
Peak transfer rates are relatively unimportant in most datacenter type situations, compared to the IOPS. People aren't generally transferring a lot of large contiguous files more so than a bunch of random reads/writes, searches, etc.
In a typical datacentre, you have machines that are generally specific to applications. Said machines often sit idle a good chunk of their life. While in a 'cloud datacentre', machines and applications are dynamically spun up or down based on actual loads. Such that there's often very high utilization of components. A single SSD, for instance, may serve as the 'boot' drive (either through iSCSI or through PXE, for instance) for literally thousands of servers.
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.
21 Comments
Back to Article
Samus - Thursday, March 31, 2016 - link
Drool. Can't wait for the consumer (a la SSD 730) model.iwod - Thursday, March 31, 2016 - link
Ok, So on a PCI-E 3.0 4x Interface the only manage to get 2GB/s. Improved Random Read....What exactly make this SSD Good when Samsung has something that is much better?
iwod - Thursday, March 31, 2016 - link
Cant Edit?? :OAnd i forgot, still 2TB only?
Billy Tallis - Friday, April 1, 2016 - link
The DC D series drives aren't going to be using 3D NAND initially, so there's no particular reason to expect a big capacity boost. The DC P3320 will eventually grow beyond 2TB, but those details haven't been announced.close - Friday, April 1, 2016 - link
As far as I can tell Samsung's only DC product at 2.5" NVMe PCIe is the PM953. It goes up to 1.92TB and the specs are 1GB/s read, 870MB/s write for the largest capacity (google MZQLV1T9HCJM).Only the PCIe cards from Samsung go up to 6TB (6GB/s read, 2GB/s write). Everything else they offer is SAS or SATA.
Or were you comparing Intel's DC SSDs to Samsung's consumer SSDs? In this case it's because you can sell anything to a consumer as long as you put a big number on the box, that you can reach only in some very specific laboratory conditions and only after you sacrifice a virgin for the speed gods. Reliability is another major difference between the two.
Otherwise you'd be pressed to ask yourself why do those guys with the datacenter constantly go for the products that have worse specs than what you can order from Amazon but many times the price...
BillyONeal - Friday, April 1, 2016 - link
Keep in mind Intel's rating here is at less than 10% variation IOP to IOP; beating the drive at 100% throughput continuously. Samsung's SM961 is rated at peak IOPs over a very short period, at least if it behaves anything like the 950 Pro. (see http://www.anandtech.com/show/9702/samsung-950-pro... and look at the IOPS graphs; and how the steady-state IOPS of Intel's 750 is around 4x what the 950 Pro manages)As a consumer drive the Samsung is probably better. As a SQL Server drive, the Intel drive is probably better. Depends on your use case.
iwod - Friday, April 1, 2016 - link
But I thought those are Random Write, Random Read doesn't seems so far off, and wouldn't in DB, higher QD gives better Random Read consistency?Samus - Saturday, April 2, 2016 - link
SQL and any server running a database (even a mail server) will have tons of tiny random writes, even in log files. The problem is since most drives don't store any user data in dram cache, and only TLC drives have an SLC "buffer" all of this random IO will wear a non-enterprise drive that lacks firmware to expect this type of environment. Enterprise drives often sacrifice a little speed by delaying buffer flushes to NAND or actually caching tiny IO data (512kb before a page write) to not hammer the NAND so much. Even Samsung has the PM841DC (or something like that) which is identical to the 840 Pro with vastly different firmware. It's also slower by like 100MB/sec!Drives that cache user data to dram cache, with the exception of many Intel drives that supposedly only use it to cache the indirection table, usually have power loss protection. Even Anandtechs review of the Intel drives suspects a little user data gets cached to DRAM though.
Kristian Vättö - Friday, April 1, 2016 - link
The SM961 is an OEM drive for the client market, so it's not comparable. Client specifications are often higher because client IO workloads are bursty, making it okay to quote maximum performance, not sustained.The drives comparable to what Intel announced yesterday are PM1725 and PM953: http://www.samsung.com/semiconductor/products/flas...
JellyRoll - Friday, April 1, 2016 - link
Aren't you an agent of Samsung now?BillyONeal - Friday, April 1, 2016 - link
Yup; was only saying that iwod's "the specs on the Samsung client drives announced are way better than this" is a flawed analysis.Kutark - Saturday, April 2, 2016 - link
Peak transfer rates are relatively unimportant in most datacenter type situations, compared to the IOPS. People aren't generally transferring a lot of large contiguous files more so than a bunch of random reads/writes, searches, etc.austinsguitar - Thursday, March 31, 2016 - link
that heatsink THOOOUGH :DDD! that looks qualitypitzel - Friday, April 1, 2016 - link
It ain't there for show. The drive can consume 30W.Samus - Saturday, April 2, 2016 - link
Pfft, it'll be fine in a ultra book. Fine!pitzel - Friday, April 1, 2016 - link
2.45A * 12V = 29.4W. That is a *lot* of power to dissipate. Why so high?Sivar - Friday, April 1, 2016 - link
As the ad says, "Bringing storage out of the cold." :)mrvco - Friday, April 1, 2016 - link
What makes a "cloud datacenter" different than a typical datacenter?smilingcrow - Friday, April 1, 2016 - link
Extra fluffyness.jwhannell - Friday, April 1, 2016 - link
A cloud datacenter is just like a normal datacenter that you run your machines in, except you don't own the hardware.pitzel - Saturday, April 2, 2016 - link
In a typical datacentre, you have machines that are generally specific to applications. Said machines often sit idle a good chunk of their life. While in a 'cloud datacentre', machines and applications are dynamically spun up or down based on actual loads. Such that there's often very high utilization of components. A single SSD, for instance, may serve as the 'boot' drive (either through iSCSI or through PXE, for instance) for literally thousands of servers.