Western Digital Announces Client NVMe SSDs With New In-House Controllersby Billy Tallis on February 28, 2018 2:00 PM EST
Western Digital has announced their first client NVMe SSDs with their SanDisk 64-layer 3D TLC NAND. These drives are also the first to feature Western Digital's new in-house NVMe SSD controllers. This is a major shift in strategy away from third-party controllers (mostly Marvell) toward complete vertical integration.
The new SSDs are called the Western Digital SN720 and Western Digital SN520. Branding for these is a bit of a mess with the drives bearing the Western Digital name and model numbers that almost fit in with the HGST Ultrastar SN200 and SN260 enterprise NVMe SSDs, but the product information is on the SanDisk website and the target market is similar to that of SanDisk's business/OEM drives like the X400 and X600 SATA SSDs. Western Digital may be trying to unify and simplify their several brands, but it's a work in progress.
The Western Digital SN720 is a high-end client NVMe SSD with a PCIe 3 x4 connection and capacities from 256GB to 1TB, with a 2TB model on the way. Aside from sequential reads, the performance of the 256GB model is a bit unimpressive, but the 512GB and larger models appear on paper to be better than the Toshiba XG5-P or Samsung 960 EVO, and close to the Samsung 960 PRO in some respects.
|Western Digital SN720 Specifications|
|Capacity||256 GB||512 GB||1 TB||2 TB|
|Form Factor||M.2 2280 Single-Sided|
|Interface||NVMe PCIe 3 x4|
|Controller||Western Digital in-house|
|NAND||SanDisk 64-layer 3D TLC|
|DRAM||DDR4, speed unknown|
|Sequential Read||3000 MB/s||3400 MB/s||3400 MB/s||TBD|
|Sequential Write||1600 MB/s||2400 MB/s||2800 MB/s||TBD|
|4KB Random Read||225k IOPS||400k IOPS||500k IOPS||TBD|
|4KB Random Write||185k IOPS||330k IOPS||400k IOPS||TBD|
|Power||Peak||9.24 W||9.24 W||9.24 W||TBD|
|PS3 Idle||70 mW||70 mW||100 mW||TBD|
|PS4 Idle||2.5 mW||2.5 mW||2.5 mW||TBD|
|Endurance||200 TB||300 TB||400 TB||500 TB|
The Western Digital SN520 is a low-end client NVMe SSD with just a two-lane PCIe link and no external DRAM. The SN520 will be available in M.2 card lengths from 30mm up to the normal 80mm, though in all sizes only the first 30mm of the PCB are actually used. The SN520 will be available in smaller capacities than the SN720, from 128GB to 512GB. Sequential throughput is limited by the PCIe x2 link, and random access performance is lower than the SN720: The 128GB SN520's random read and write performance are about what SATA SSDs advertise. The 256GB SN520 is very close to the random access performance of the 256GB SN720, but at 512GB the SN520 is well behind the SN720.
In spite of its limitations, the SN520 may end up beating the Western Digital WD Black NVMe SSD on most benchmarks, thanks to the better performance of SanDisk's 64L 3D TLC as compared with the 15nm planar TLC used by the WD Black.
The Western Digital SN720 and SN520 both have five-year warranties and the same write endurance rating at each capacity. The 128GB model is rated for 100 TB (0.42 DWPD), but each doubling of drive capacity only brings an additive increase of 100 TB to the write endurance rating, leaving the 1TB SN720 with just 400 TB (0.21 DWPD). The smaller capacities have great write endurance ratings by the standards of consumer TLC drives, but the largest capacities definitely don't have premium endurance ratings.
|Western Digital SN520 Specifications|
|Capacity||128 GB||256 GB||512 GB|
|Form Factor||M.2 2280, 2242, 2230 Single-Sided|
|Interface||NVMe PCIe 3 x2|
|Controller||Western Digital in-house|
|NAND||SanDisk 64-layer 3D TLC|
|DRAM||None (Host Memory Buffer not supported)|
|Sequential Read||1500 MB/s||1700 MB/s||1700 MB/s|
|Sequential Write||800 MB/s||1300 MB/s||1300 MB/s|
|4KB Random Read||95k IOPS||220k IOPS||270k IOPS|
|4KB Random Write||90k IOPS||175k IOPS||280k IOPS|
|Power||Peak||3.96 W||4.62 W||5.94 W|
|PS3 Idle||25 mW||25 mW||25 mW|
|PS4 Idle||2.5 mW||2.5 mW||2.5 mW|
|Endurance||100 TB||200 TB||300 TB|
At the heart of these new products are Western Digital's first in-house NVMe SSD controllers. Western Digital isn't providing much information about the shared controller architecture used by these products, but we have a bit of information. The controllers are manufactured on a 28nm process, which is the same node that most other high-end NVMe controllers for the consumer market uses. The controllers support 8 channels for NAND access and have a DDR4 DRAM controller. The choice of DDR4 instead of LPDDR4 is probably more cost-effective for Western Digital since they do not manufacture their own DRAM, but it may also hurt the power efficiency of the SN720 somewhat: LPDDR4 seems to be a major factor in the big jump in efficiency Samsung's 860 PRO and 860 EVO show over their 850 generation.
The SN520 doesn't have any external DRAM and Western Digital won't say what kind or quantity of memory is in the controller package itself, but there's a good chance the SN520 has its DRAM controller deactivated. The SN520 may also be operating with just 4 NAND channels, similar to other low-end NVMe controllers like the Silicon Motion SM2263 and Phison E8. Western Digital did confirm that the SN520 does not use the NVMe Host Memory Buffer feature that alleviates the worst performance problems of DRAMless SSDs, but if there's some on-controller memory then HMB isn't necessary.
Western Digital hasn't disclosed what kind of processor cores are used in their NVMe controllers, but they did confirm that these aren't using the RISC-V architecture—those products won't be arriving until next year at the earliest. The Western Digital NVMe controllers are probably using ARM Cortex-R cores like most SSD controllers.
The Western Digital SN720 and SN520 are currently sampling to select customers. The time frame for volume availability has not been announced. Until now, Western Digital has taken a very reluctant approach to NVMe for non-server applications. With the launch of their in-house controllers, this is sure to change. These two tiers of client NVMe SSDs for the business/OEM markets will give Western Digital broad coverage of those markets, lacking only a fully-integrated BGA SSD solution. We expect consumer-oriented derivatives of at least one of these two products to come to the retail market this year, but Western Digital hasn't said anything official about any such plans.
Post Your CommentPlease log in or sign up to comment.
View All Comments
Mikewind Dale - Wednesday, February 28, 2018 - linkMr. Tallis,
Speaking of RAM, you've raised a more general question in my mind, and I'd like to see an Anandtech article about it:
What capabilities do different brands and types of SSDs have to mitigate power-loss? From what I can tell, consumer SSDs have been moving away from having power-loss capacitors - especially because the M.2 form factor simply lacks the space.
And to what extent does having a battery or UPS effectively substitute for the lack of power-loss capacitors?
And what if the system suffers, not a power-loss, but another form of system crash? If the system crashes, but the battery or UPS still supplies power, will an SSD be able to write all of the extent RAM data to NAND?
Intuitively, I would think that if the SSD has power-loss capacitors, it will be more robust because it isn't relying on power being continuously fed through the motherboard from a battery or UPS. So even if the rest of the system crashes in some way, the SSD controller still has its own dedicated source of power. But that's just my intuitive guess. I have no idea whether that's right.
And does it make a difference whether the RAM is on the SSD or whether it uses HMB (system memory)? Is the SSD more capable of finalizing all pending transactions and writing them to NAND, depending on whether the RAM is local or system?
In general, I'd like to see a discussion of how power loss and system crashes affect SSDs, and whether any brands or types of SSDs are more robust than others.
Personally, I'd prefer a greater degree of robustness assuming the cost in money or performance isn't too high. Money is obviously a major factor. There's a reason I'm not buying a Xeon with ECC RAM and an enterprise SSD. Money talks! But if it is possible to obtain a greater degree of robustness without sacrificing too much performance or money, I'd love to know how.
Billy Tallis - Wednesday, February 28, 2018 - linkConsumer SSDs generally don't use their DRAM as a write cache for user data. It's a cache for the drive's Flash Translation Layer (FTL) that maps logical block addresses (LBAs) to phyiscal flash pages and blocks. The FTL is structured very much like a journalling filesystem, so it has some resilience, and updates to the mapping tables are written out to the flash fairly quickly. NVMe drives using HMB can only treat that memory as a write-through cache and must be prepared for the sudden loss of access to that memory. That's a small part of why HMB isn't a complete substitute for on-drive DRAM.
If your system crashes but the drive is still getting power, then it can finish flushing its write caches (which are typically small amounts of SRAM on the controller, plus the page buffers on the NAND chips themselves). Even if the drive does suddenly lose power, it's pretty hard for the FTL to be corrupted—you just lose some of the most-recently written data. This is where the Crucial MX series has a slight advantage, because they include enough capacitors to ensure that a write that can't be completed doesn't corrupt data that was already written to the flash. Since all modern operating systems use journalling filesystems, they are also tolerant of a bit of data loss.
The OS can signal to the drive when a write command must be sent directly to the flash and not be stored in a volatile buffer. The filesystem will do this at appropriate times when updating critical data structures and database applications do this to ensure transaction consistency, but most consumer applications have no reason to issue their writes in this manner. Drives can ignore these signals, and enterprise SSDs with full power loss protection are free to do so without risking data loss. All the NVMe SSDs I've tested seem to respect the commands to flush their write buffers or perform a write without buffering, and that's a problem because the Windows NVMe drive defaults to performing that kind of write way too often.
It's worth noting that every SATA SSD I test gets hot-swapped at least a dozen times. The drives aren't actively being accessed by the host system when they're ejected, but in many cases the drives are likely to still be performing background operations. This hardly ever kills the drives, and no model I've tested has ever shown systemic problems with hot-swapping. (I'm not currently checking data integrity after hot-swaps, though.)
WinterCharm - Wednesday, February 28, 2018 - linkVery insightful. Thanks for taking the time to write that out :)
Mikewind Dale - Wednesday, February 28, 2018 - linkThanks!!
MajGenRelativity - Thursday, March 1, 2018 - linkThank you very much for that detailed response! It's information like this that keeps me coming back to Anandtech. Whenever I want all the information I can handle, I come here
Sivar - Tuesday, March 6, 2018 - linkThis forum reply alone is better than most entire articles on other tech sites.
XabanakFanatik - Wednesday, February 28, 2018 - linkI'm not sure where you get "and close to the Samsung 960 PRO in some respects", the sequential reads are very close, while the sequential write and random read and write exceed the paper specifications for the 960 Pro series. 3/4 paper performance specs is not "in some respects".
I am extremely interested to see the performance comparison because of this.
Billy Tallis - Wednesday, February 28, 2018 - linkWithout a lot more details from Western Digital about the test conditions their specs are derived from, there isn't enough information to justify any kind of confident claim that the SN720 will be competitive with the performance of the 960 PRO. We know right off the bat that the SN720's peak write performance will be limited by its SLC cache, while the 960 PRO doesn't use SLC caching.
XabanakFanatik - Wednesday, February 28, 2018 - linkYou have a good point, I might be getting a little ahead of myself. I'm really excited to see an opportunity for some competition in the high end consumer NVM-e space. I hope it pans out.
jjj - Wednesday, February 28, 2018 - linkInteresting that they have their own controller. Pliant did have their controller and maybe the same team developed this one.Hard to be certain as WD seems to intend to invest a lot more in their own chips.
Long term that makes sense if the future brings processing closer and closer to memory and then, the 2 merge with brain like devices.