Random Read Performance

Our first test of random read performance uses very short bursts of operations issued one at a time with no queuing. The drives are given enough idle time between bursts to yield an overall duty cycle of 20%, so thermal throttling is impossible. Each burst consists of a total of 32MB of 4kB random reads, from a 16GB span of the disk. The total data read is 1GB.

Burst 4kB Random Read (Queue Depth 1)

The Optane SSD 900P doesn't break the record for QD1 random reads, but only because we've also tested the 32GB Optane Memory M.2, which is about two microseconds faster on average for each 4kB read. The Optane SSD 900P is still about 7 times faster than any flash-based SSD.

Our sustained random read performance is similar to the random read test from our 2015 test suite: queue depths from 1 to 32 are tested, and the average performance and power efficiency across QD1, QD2 and QD4 are reported as the primary scores. Each queue depth is tested for one minute or 32GB of data transferred, whichever is shorter. After each queue depth is tested, the drive is given up to one minute to cool off so that the higher queue depths are unlikely to be affected by accumulated heat build-up. The individual read operations are again 4kB, and cover a 64GB span of the drive.

Sustained 4kB Random Read

When longer transfers and higher queue depths come into play, the Optane SSD 900P passes the Optane Memory M.2 and remains more than 6 times faster for random reads than any flash-based SSD.

Both Optane devices more or less level off at queue depths of 8 or higher. The Optane SSD 900P saturates at about 1800 MB/s while the Optane Memory tops out around 1300 MB/s. The Samsung 960 PRO 2TB hasn't caught up by QD32, and doesn't surpass the QD1 random read performance of the Optane SSD until the Samsung reaches a queue depth of about 8.

Random Write Performance

Our test of random write burst performance is structured similarly to the random read burst test, but each burst is only 4MB and the total test length is 128MB. The 4kB random write operations are distributed over a 16GB span of the drive, and the operations are issued one at a time with no queuing.

Burst 4kB Random Write (Queue Depth 1)

The burst random write performance of the Optane SSD 900P is slightly higher than the Intel SSD 750 1.2TB, and about 14% faster than Samsung's fastest.

As with the sustained random read test, our sustained 4kB random write test runs for up to one minute or 32GB per queue depth, covering a 64GB span of the drive and giving the drive up to 1 minute of idle time between queue depths to allow for write caches to be flushed and for the drive to cool down.

Sustained 4kB Random Write

With higher queue depths in play, the Optane SSD 900P scales up faster than the Intel SSD 750 1.2TB, leaving the Optane SSD with a 7-10% lead over the Samsung 960s and Intel 750.

Samsung's 960 PROs and the larger 960 EVO all trail slightly behind the Optane SSD's random write performance for queue depths 1 to 4, then the Samsung drives level off and leave the Optane SSD with a substantial performance advantage at high queue depths. The Intel 750 is slightly faster at QD1 and QD2, but saturates at an even lower performance level than the Samsung 960s.

AnandTech Storage Bench - Light Sequential Performance
POST A COMMENT

205 Comments

View All Comments

  • ddriver - Friday, October 27, 2017 - link

    I got myself a bucked of salt. The necessary requirement to swallow that Houdini "2.7x better" claim from the launch PR.

    I've been rendering stuff since the days of 3d max for frigging DOS. And I am yet to experience a scenario where CPU load is not in the 99% range.

    Having a rendering job that cannot feed the CPU to above 10% load with the insanely fast 960 pro has got to be an unprecedented case of cooked-up benchmark in human history.
    Reply
  • extide - Friday, October 27, 2017 - link

    Did you read the article? It pretty clearly explains how they got that result, and it makes sense. Reply
  • ddriver - Friday, October 27, 2017 - link

    Oh yeah, I get it. Hypetane is a synthetic beast. Which allows to showcase said advantage as long as you focus on it in a carefully devised and completely detached from real-world usage workload.

    Don't get me wrong. It is good that hypetane is now available in capacities that actually allow to use it. And if endurance turns out to be tangibly better than nand, I might actually buy it. Low queue depth performance is good, especially random read, which may not be of that much practical use to most of the people out there, but I could make good use of that.

    But it will remain "hypetane" even after I go and buy it. Because intel said "1000 times better", and it is not even 10 times better. A zero on its own might be nothing, but two zeroes after a positive number make quite a lot of difference.
    Reply
  • ddriver - Friday, October 27, 2017 - link

    "no other alternative nonvolatile memory technology is close to being ready to challenge 3D XPoint"

    Except for SLC, which was so good it was immediately abandoned once inferior and more profit friendly NAND implementations were available.

    A SLC based product coupled with MRAM cache will easily humiliate hypetane in its few strong aspects.

    Too bad NAND drives are now moving to TLC and QLC, even MLC is heading in the "luxury item" category. Too bad because 3D SLC has tremendous potential. Let's see if it gets realized.
    Reply
  • extide - Friday, October 27, 2017 - link

    How would that work. SLC is slower than Optane, can't be written at a block level, needs trash collection, etc. Then you cache it with a technology similar to Optane? Why not just build a drive with all MRAM, oh yeah, too expensive. Looks like Optane wins. Reply
  • ddriver - Friday, October 27, 2017 - link

    Nope, SLC is actually faster. Look it up.

    And what it cannot do is write at the bit level. Which is not really a big deal. Even CPUs cannot address RAM at bellow a byte, if you want single bit operations, you have to use bitwise operators. Writing at a higher level is actually very efficient, because it reduces overhead. If single bit addressing was important, that's who computers would work.

    Furthermore, single bit writes produce a significant challenge when tracking wear levels. Hypetane still wears out, you know... It will be tremendously harder to accurately track wear at bit level, and I am abot 99.999999% sure it is not how intel does it, meaning that a lot of that supposed extra endurance will be forfeited by managing wear at a coarsely grained level. They won't be managing that at bit level, the overhead will be tremendous and will completely diminish potential advantages.

    The MRAM cache will reduce a lot of write amplification and garbage collection.

    It also looks like 3d SLC has about 3 times the density of the chips intel is currently using for hypetane.

    "Why not just build a drive with all MRAM" - density is too low. Which is also why we use RAM for working memory, I mean volatility can easily be solved by say adding a RTG battery to a DRAM drive, giving it effectively about a century of continuous, uninterrupted power. It is doable, but then again, redundant, and while it is true that the industry does a lot of pointless things nowadays, the only ones that qualify are those with a desirable usability to profitability ratio, and a RTG DRAM drive is simply too good to offer...

    "Looks like Optane wins" - anyone can win when running unopposed. The moment someone makes a SLC/MRAM hybrid and it loses to hypetane, I will retract my statement and admit I was wrong. I have zero problem with that ;)
    Reply
  • vanilla_gorilla - Friday, October 27, 2017 - link

    So you're saying Optane sucks because it would be slower than a drive that doesn't exist? Reply
  • ddriver - Friday, October 27, 2017 - link

    No, I am saying it "sucks" because for all intents and purposes, it is not any faster than a 2 year old drive that it was supposed to beat by a 1000 times.

    And the reason I put it "sucks" is because I never said it does suck. I give it a very realistic valuation. What sucks is how far that realistic valuation is from what intel promised. Which is entirely on them.
    Reply
  • name99 - Friday, October 27, 2017 - link

    He's saying two distinct things.
    (a) This costs too much for what it delivers. IF Samsung wanted to compete with it, they could do so with a suite of existing technologies. But they probably won't do so because there is little demand for a product like this; honestly it only exists so that Intel can say "see, 3D-XPoint is too, real".

    (b) The place where 3D-XPoint ACTUALLY makes sense is, more or less, what AnandTech says --- as a slower (but much larger) RAM replacement. That's what plays to the technology's strengths (simple controller, byte-level access). But Intel STILL are not shipping that --- which makes one wonder WTF not?

    It IS reasonable to point out that Intel has been lying about this product since the day it was announced, and that the only reason they're shipping these SSD drives is to throw up more smoke to hide the fact that the actually sensible use case remains (for some reason) impossible.

    Being a fanboy isn't about always praising your company, it's about refusing to criticize your company even when they're clearly in the wrong. Intel is clearly in the wrong here, in the sense that nothing that they promised about Optane is actually reality even today, two years after the announcement.
    If you think that's reasonable behavior, ask yourself how you would react if your favorite villainous company did the same.
    Would you be impressed if AMD announced that they're going to ship a GPU 1000x faster than the competition, and two years later all they have is something 2.7x as fast (under very specialized circumstances)?
    Would you let Apple off the hook if they said that the Apple car was going to have 1000x the range of a Tesla, then they shipped two years later, a car with 2.7x the range of a Tesla?
    Reply
  • Drumsticks - Saturday, October 28, 2017 - link

    Re: AMD example: if AMD claimed a product would be 100x or 1000x faster than Nvidia, but only delivered something 6-10x faster in the majority of cases, and on par in the rest, for only 2-3x more money, I'd still be pretty satisfied. Reply

Log in

Don't have an account? Sign up now