Burst IO Performance

Our burst IO tests operate at queue depth 1 and perform several short data transfers interspersed with idle time. The random read and write tests consist of 32 bursts of up to 64MB each. The sequential read and write tests use eight bursts of up to 128MB each. For more details, please see the overview of our 2021 Consumer SSD Benchmark Suite.

QD1 Burst IO Performance
Random Read Random Write
Sequential Read Sequential Write

The WD Black SN850 turns in excellent scores on almost all of the burst IO tests. For random reads, it edges out the Intel SSD 670p to set a new record for flash-based SSDs, and even when testing beyond the bounds of any possible SLC caching it is only 2% slower than the MLC-based Samsung 970 PRO. For random writes the WD Black SN850 is slightly slower than the Phison E16 drive, but otherwise is s clear step up in performance from the rest of the field. When testing sequential transfers on a small slice of the drive, the SN850 is substantially faster than everything else, but when testing across 80% of the drive its sequential read performance drops dramatically and is beat by the Samsung 980 PRO and several of the faster PCIe 3.0 drives.

Sustained IO Performance

Our sustained IO tests exercise a range of queue depths and transfer more data than the burst IO tests, but still have limits to keep the duration somewhat realistic. The primary scores we report are focused on the low queue depths that make up the bulk of consumer storage workloads. For more details, please see the overview of our 2021 Consumer SSD Benchmark Suite.

Sustained IO Performance
Random Read Throughput Power Efficiency
Random Write Throughput Power Efficiency
Sequential Read Throughput Power Efficiency
Sequential Write Throughput Power Efficiency

On the longer random read test, the WD Black SN850 doesn't quite stand out from the best performance offered by other drives with newer flash. But on the other three workloads the SN850 is clearly superior, with significant performance leads over the rest of the competition. Its power consumption is consistently on the high side and in some cases it is drawing more than any of the other drives, but the performance is high enough that the efficiency scores are all good.

Random Read
Random Write
Sequential Read
Sequential Write

For random reads, the SN850 eventually ramps up to around 4GB/s or 1M IOPS at the end of the test, which is significantly faster than any other drive that we've tested so far on this new test suite. However, when testing across 80% of the drive instead of just a 32GB slice, the random read performance falls to roughly the same level as the Samsung 980 PRO.

For random writes, the SN850's performance scales up a bit quicker than the 980 PRO, but it hits a throughput limit sooner and the 980 PRO ends up being much faster for random writes to the SLC cache at high queue depth.

For sequential reads, the SN850 ends up slightly faster than the 980 PRO, but when testing across 80% of the drive the Samsung reaches full performance with a lower queue depth. For sequential writes the SN850 is again a bit faster than the 980 PRO and this time it doesn't need higher queue depths to reach full speed, but it also starts running out of SLC cache before the test is over while the 980 PRO maintains full performance through the end of the test.

Random Read Latency

This test illustrates how drives with higher throughput don't always offer better IO latency and Quality of Service (QoS), and that latency often gets much worse when a drive is pushed to its limits. This test is more intense than real-world consumer workloads and the results can be a bit noisy, but large differences that show up clearly on a log scale plot are meaningful. For more details, please see the overview of our 2021 Consumer SSD Benchmark Suite.

The WD Black SN850 starts off this test with good random read latency, but around 80k IOPS it shifts gears and latency spikes alarmingly. It actually improves a few times later in the test so by the time the drive is approaching its throughput limit, it is only a bit slower than the Samsung 980 PRO.

Trace Tests: AnandTech Storage Bench and PCMark 10 Advanced Synthetic Tests: Block Sizes and Cache Size Effects
POST A COMMENT

82 Comments

View All Comments

  • Pinn - Thursday, March 18, 2021 - link

    Would love to see thermals. Reply
  • JoeDuarte - Thursday, March 18, 2021 - link

    Why are the write latencies so much lower than the read latencies? (For all the drives.) Is this normal for SSDs? I hadn't noticed this pattern before, or read anything about it. My assumption is that reading should be faster than writing.

    To really move the needle on latency we'll need to move away from PCIe to something like OpenCAPI, which is a much faster interface. Optane can't really stretch out to its full potential if it's going to be hitched to PCIe, even 4.0. With the end of Moore's Law, we really need to optimize the I/O as much as possible, and get rid of interfaces and buses that require many thousands of CPU cycles per transaction.

    By the way, why is there no energy usage data for the Optane drive in the results? It seems to be missing for all benchmarks. That drive is in all the performance results except energy usage.
    Reply
  • Billy Tallis - Thursday, March 18, 2021 - link

    Reading a single page from NAND flash is a lot faster than programming a page. But writes can be cached and several smaller writes can be saved up to be issued in a batch that better uses the parallelism inside the SSD. So the amortized cost of writes can be much lower. Of course, this poses some risk to data in the event of power loss, but that's a generally-accepted tradeoff for consumer systems.

    The power data for the Optane 905P was left off because its idle power is higher than the peak load power of almost all of the other drives. There aren't a lot of interesting comparisons to be made there. The Optane drive is always the most power-hungry, by far. It would be even without the RGB LEDs. It only has a chance of being competitive on power efficiency for low-QD random reads.
    Reply
  • Kamen Rider Blade - Friday, March 19, 2021 - link

    Optane is so perfect as a home DeskTop OS drive where the low QD and latency really can be taken advantage of along with it's Random IO and Latency advantages.

    The vast majority of home users are 90/10 Read/Write.
    Reply
  • Spunjji - Friday, March 19, 2021 - link

    Only, for that usage, the price/capacity trade-off makes it poor value for money - and the advantages it does confer are barely noticeable in use. Reply
  • FunBunny2 - Friday, March 19, 2021 - link

    "Optane is so perfect as a home DeskTop OS drive where the low QD and latency really can be taken advantage of"

    I would argue the opposite: Optane, et al, make the most sense for industrial strength RDBMS, used in App Mode.
    Reply
  • Oxford Guy - Sunday, March 21, 2021 - link

    'Of course, this poses some risk to data in the event of power loss, but that's a generally-accepted tradeoff for consumer systems.'

    Didn't some consumer SSDs have a capacitor to prevent data loss? Has that feature been lost due to the smaller form factor (versus SATA), or is it mainly due to cost-cutting?
    Reply
  • Billy Tallis - Sunday, March 21, 2021 - link

    There may have been a few "consumer" SSDs back in the very early days that had full power loss protection, but that has been an enterprise-only feature for as long as SSDs have been even remotely mainstream for consumers. (Exceptions: Intel 750 and Optane SSDs, which are re-branded enterprise drives and do have power loss protection.)

    There have been some consumer SSDs with partial power loss protection, designed to prevent data already on the drive from being corrupted by later writes that get interrupted by a power loss (but making no guarantees about completing any in-progress writes). This doesn't require extra capacitors for writes to SLC or any other single-pass writing (which includes a lot of TLC, if not all of it these days). And since there are also other good reasons not to leave a page in a partially-programmed state for long, I suspect most consumer SSDs have moved away from ever needing the kind of capacitor banks we saw on eg. early Crucial MX series drives.
    Reply
  • Oxford Guy - Monday, March 22, 2021 - link

    I can imagine that consumers who spends thousands on things like GPUs would be hard-pressed to pay for a capacitor.

    Good thing the flash drive companies are so watchful of our crucial pennies.
    Reply
  • Mikewind Dale - Friday, March 26, 2021 - link

    Does having a laptop battery or desktop UPS effectively take the place of power-loss capacitors on an SSD? I would think it does, but I'd like to be sure. Reply

Log in

Don't have an account? Sign up now