AnandTech Storage Bench - The Destroyer

Our AnandTech Storage Bench tests are traces (recordings) of real-world IO patterns that are replayed onto the drives under test. The Destroyer is the longest and most difficult phase of our consumer SSD test suite. For more details, please see the overview of our 2021 Consumer SSD Benchmark Suite.

ATSB The Destroyer
Average Data Rate
Average Latency Average Read Latency Average Write Latency
99th Percentile Latency 99th Percentile Read Latency 99th Percentile Write Latency
Energy Usage

For SATA drives, the Samsung 870 EVOs turn in class-leading scores on almost all of the performance metrics. But these improvements are all marginal at best; the SATA interface bottleneck almost completely levels the playing field. The small improvements to read latency brought by the 870 EVO pale in comparison to what is achieved by even entry-level NVMe SSDs.

In stark contrast to the performance numbers, the 870 EVOs turn out to be the most power-hungry TLC drives in this bunch: they sacrifice some of the efficiency improvements the 860 EVO provided, even though drives like the SK hynix Gold S31 have been able to deliver significant improvement on this.

AnandTech Storage Bench - Heavy

The ATSB Heavy test is much shorter overall than The Destroyer, but is still fairly write-intensive. We run this test twice: first on a mostly-empty drive, and again on a completely full drive to show the worst-case performance.

ATSB Heavy
Average Data Rate
Average Latency Average Read Latency Average Write Latency
99th Percentile Latency 99th Percentile Read Latency 99th Percentile Write Latency
Energy Usage

The scores for the Heavy test paint much the same picture as for The Destroyer. The full-drive test runs additionally show that the worst-case performance of the mainstream SATA SSDs is still superior to many entry-level NVMe SSDs, even though the NVMe SSDs significantly outperform SATA for any more normal workload.

AnandTech Storage Bench - Light

The ATSB Light test represents ordinary everyday usage that doesn't put much strain on a SSD. Low queue depths, short bursts of IO and a short overall test duration mean this should be easy for any SSD. But running it a second time on a full drive shows how even storage-light workloads can be affected by SSD performance degradation.

ATSB Light
Average Data Rate
Average Latency Average Read Latency Average Write Latency
99th Percentile Latency 99th Percentile Read Latency 99th Percentile Write Latency
Energy Usage

On the Light test, the measurable but imperceptible performance advantages of the 870 EVOs over other SATA drives have basically disappeared. The read latency scores on the full-drive test runs may be a tiny bit better than the 860 EVO, but the only scores that have clearly shifted with this new generation are the energy consumption figures that have creeped up.

PCMark 10 Storage Benchmarks

The PCMark 10 Storage benchmarks are IO trace based tests similar to our own ATSB tests. For more details, please see the overview of our 2021 Consumer SSD Benchmark Suite.

PCMark 10 Storage Traces
Full System Drive Overall Score Average Bandwidth Average Latency
Quick System Drive Overall Score Average Bandwidth Average Latency
Data Drive Overall Score Average Bandwidth Average Latency

The Full System Drive test from the PCMark 10 Storage suite shows a much wider spread of performance scores among SATA drives than our ATSB traces, but also a much smaller advantage for the NVMe drives. Judging by this test, the 870 EVO offers a small but real improvement to performance compared to earlier SATA drives. The 4TB 870 QVO also scores quite well since it benefits from the same controller and has enough SLC cache to almost match the performance of the 4TB 870 EVO.

The subset of tests included in the Quick System Drive and Data Drive benchmarks show a more level playing field among SATA SSDs, and a greater advantage for NVMe drives. Since we run these tests before the Full System Drive test, each drive is closer to its fresh out-of-the-box state, which helps these tests get closer to showing the theoretical peak performance of a drive.

Introduction Synthetic Tests: Basic IO Patterns
Comments Locked


View All Comments

  • hansmuff - Wednesday, February 17, 2021 - link

    Especially the 4TB seems like a fantastic Games drive to me, really good performance at a great price.
  • ekon - Wednesday, February 17, 2021 - link

    What (my) world needs is an absolutely rubbish but cheap high capacity SSD. As many cell levels as it takes.
  • jarablue - Wednesday, February 17, 2021 - link

    I have 2 sata WD Blue 1 tb ssds for game installs. They work totally fine and load games fast as hell. SATA ssds are still on point for large game storage space.
  • Spunjji - Friday, February 19, 2021 - link

    It'll be interesting to see if this changes along with software being developed for the new generation of consoles.
  • Duncan Macdonald - Thursday, February 18, 2021 - link

    It takes 10 gig ethernet to exceed the speed of SATA - the SATA limit of 600 MBytes/sec is 4800Mbits/sec - allowing for TCP/IP overhead even a 5GbE link can not carry data as fast as a SATA link.

    SATA is still the only effective way of increasing the internal storage of systems that have no free NVMe slots available.
  • Tomatotech - Thursday, February 18, 2021 - link

    Or a USB3 / USB-C drive taped / velcro’d somewhere in the PC case.

    Could contain either 2.5” HDD, or 2.5” SDD or a m.2 NVME SSD in a USB enclosure. I’ve done that a couple of times with small cases. Works perfectly fine especially with solid state media.
  • edzieba - Thursday, February 18, 2021 - link

    Cable-attached storage still has the massive advantage that you can connect more drives than you have board area for. A regular motherboard may have one m.2 slot, paying out the nose may net you two slots. Selling off a few organs may buy a halo board with 3 or more slots. Or you may need a bloated riser card and occupy as 16x slot (no go for ITX). Or... you can use the at least 4 SATA ports on even the most bargain basement board (with 8 being hardly uncommon) to stuff more capacity in as needed with ease.
    SATA Express was the right idea at the wrong time: a x1 or x2 PCIe interface to allow NVMe, with PHY fallback to SATA, and at entirely acceptable bandwidth for most uses (stick in an m.2 boot drive for OS and key applications) would be a perfect upgrade path for consumer SATA use. If it integrated power transport too (for SSD only, block mechanical drives with keying at the device end cable) it would simplify cable routing too. But that boat has probably sailed for good, unless enterprise just happens to adopt such a connector and drive architecture, which seems unlikely with density demands ever increasing. I don't think we'll see any new internal high bandwidth cabling standards other than PCIe link rate updates to the persistently high-priced OCuLink.
  • Tomatotech - Thursday, February 18, 2021 - link

    For more internal storage, a USB3 / USB-C drive can be taped / velcro’d somewhere in the PC case.

    Could contain either 2.5” HDD, or 2.5” SDD or a m.2 NVME SSD in a USB enclosure. I’ve done that a couple of times with small cases. Works perfectly fine especially with solid state media.
  • abufrejoval - Thursday, February 18, 2021 - link

    I remember experimenting with Compact Flash cards on PCMCIA and IDE adapters, trying to run Windows XP and Linux on them: Sure, there were no seeks, but at the time I didn’t understand the erase block issues yet and was just befuddled how some I/O just seemed so slow it had XP crash.

    When FusionIO came out with their first devices, I jumped on those and they were basically a precursor of NVMe, ouch, is it 13 years already?

    I’ve celebrated SATA SSDs, still have a 160GB Postville under current in a firewall, that may last another 10 years. I lost count, but there may be 30-50 SATA SSD in the house, some still used as „boot stick“ with only 128GB, most 1/4 to 1TB, some with 2TB.

    The only 4TB SSD here is actually a RAID-0 using 4x1TB, because I tend to have plenty of SATA ports left over in all these tower chassis, that used to house 3.5“ HDDs. The last system with 2.4 TB FusionIO card also still sports 4x 200GB Intel enterprise DC3700 MLC drives, just because that X99 board has 10 SATA ports, so why not use them in a RAID so vastly overprovisioned that it will never die?

    Like those ancient HDDs, these SSDs move around between boxes, almost like the „Winchesters“ or removable hard disks in the old days (been around since PDP-11/34). I use carrier-less hot swap bays on all systems, SATA caddies on laptops or USB3/SATA cases, for their flexibility: Milliseconds saved on storage benchmarks don’t compare to productivty and not having to disassemble a workstation with a 3 slot GPU and giant fans for low-noise, just to switch to another OS is a real bonus!

    Most of the time SATA-SSD really is quite simply fast enough. If not, it’s the architecture, stupid!

    And some of the more agonizing waits, turn out to be not at all related to SATA vs. NVMe…

    ARK Survival Evolved is one of my favorite games, because I play that with my kids. Its main downside is loading-time: It just takes ages and ages to launch! Sure, it has 200GB data with all these extra maps and extensions, but perhaps more importantly, it’s files.

    I got really tired of waiting for those minutes it took to load that from HDDs, so I invested in one of those „giant“ 1TB (SATA) SSDs at the time… still took awhile to load, but the improvement was significant (less than one CMU or coffee mug unit). Now, since we all play the game, I tried to be smart and put it on the network in the 4TB JBOD/RAID0 I mentioned, and then upgraded to a 10Gbit network to match the performance.

    Alas, the load times across the 10Gbit network were horrific! Far worse, than the single 2TB HDD I had used in the beginning.

    Then one day I ran ARK on Linux, within a larger experiment on the quality of Linux gaming. I didn’t have a big enough SSD around to store the game data, so instead I used one of those 2TB -WD HDD hunks from 15 years back, that just refuse to fail.

    And then I almost fell of my chair, when ARK launched faster off that HDD than I had ever even see it launch from an NVMe drive (yes, of course, I had to have some of those, too).

    Long story slightly abbreviated, the annoyingly slow ARK launches were never a storage issue, but a Windows file access overhead issue. Linux truly put Windows to shame that day! It managed to load those tens of thousands of files ARK required much faster from a HDD, than Windows needed on NVMe storage, and way, way, WAY faster than Windows (10Gbit) networking from a SATA-SSD RAID0.

    Now, Windows and Windows 10Gbit networking isn’t always and by-default orders of magnitude slower than Linux. At least not, when you’re dealing with a few large files. But when your game (or application) happens to use small files instead of 10 big ones, be advised to test the OS before you blame it on the storage!

    The general protocol and latency overhead of SATA vs. (PCIe)NVMe is no doubt significant.
    As are the benefits of a well established form factor with all those ports and enclosures I already own, and the flexibilty I learned to rely on. Dogma rarely helps and I find myself buying SATA-SSD over NVMe, once the default system boot storage requirements for every box have already been filled (with NVMe, if capable). Mostly because a) SATA SSD are really fast enough already (real bottlenecks are architecture) and b) because aggregating those lower capacity NVMe sticks into RAID0 to extend their usability, is really, really, really expensive, at least so far, because those PCIe switches are so overpriced by Avago/Broadcomm, while SATA multiplexers are cheap and mostly built-into the PCH you already own.
  • zodiacfml - Thursday, February 18, 2021 - link

    I don't mind SATA being limited since it is only huge file transfers are limited, not random performance or game/OS loading. The form factor needs improving though. A quick, cheap way to do that is to simply cut the 2.5" form factor to half or smaller since it will still leave a pair of mounting holes for screws.

Log in

Don't have an account? Sign up now