Optane 1600X 118GB - Lots of CDM benching and some thoughts

[QcRMoD]

These CDM benchmarks of the P1600X were made on a clean Win 11 Pro 23H2 install for the first run. Then as a second step, to sort of parallel the conditions in which W1zzard tests these, I filled the drive furthermore to about 80% capacity to rerun all tests – twice.

It’s not like I went into the fresh install for that. In fact, the main goal of the operation was to install Windows 11 on an ineligible system AND live with it for a few days. But as I did so I thought I should bench the drive, since there is so little verifiable benchmarks of these out there.

Core system:

• Maximus IX Code
• I7 7700K
• 2x8GB DDR4 3200Mhz
• BIOS is per default except XMP profile applied and HD Audio/WiFi/Bluetooth disabled.

So I guess there could be better tests from someone with newer and better hardware that also adheres to what Win 11 considers worthy of its incorrigibly extensive collection of personal data and money... aaah so sorry, what am I saying... it’s for security reasons, obviously!

Still the P1600X is running on par to its specs on the main PCIe 3.0x4 M.2 slot from the PCH. Little done to Win 11 except installing a few everyday programs and cutting the prying to an acceptable minimum. No drive optimization nor specific drivers other than what Win 11 deemed proper.

Last thing I would like to add is that I had not used CDM for benching anything since the old days when Sata SSD changed our world and it became all the hype among enthusiasts to use CDM and show their results. Now there’s a nVME SSD setting that did not exist back then, and Peak/Real World profiles that might have but I couldn’t remember. Anyhow I got carried on trying each as-is combination, as I wanted to cover those tests often used in shady marketing tactics to hype the numbers. That nVME setting seems explicitly intended for that.

Strictly speaking, one would only need to run one pass with Real World settings, one pass with Default/Peak settings, then one pass with Nvme SSD/Peak Performance settings switching SEQ1M Q8T1 to SEQ128K Q32T1, to cover the same ground I did here. No test wasted in redundancy, and voilà. It even has the beauty to keep Q1T1 stuff together AND hype stuff together. I kept with redundancy for this post however because it gave 9/12 samples RND4K Q1T1/SEQ1M Q8T1, which are after all the strict numbers we can look at to get the real idea of the drive’s peak performance and of its “OS” capabilities.

Here’s a compilation I made in Excel averaging all tests. I even left there for you the RND4K Q1T1 data set I compiled last.

I got a lot of variance with SEQ128K and RND4K Q32T16 writes, which clearly usually aim to give the highest results. In the Compilation SEQ128K in fact has the lowest results of Write and Mix even compared to SEQ1M Q1T1. You’ll see a good example of such variance in the Verification results in the last post: MUCH better SEQ128K writes there but conversely quite lower RND4K Q32T16 writes to still a better Mix score. I have no explanation to offer about that much – it might be typical test behavior.

I take note that the more we circle closer to Q1T1 either Sequential on Random, the less the Optane gets any variance. SEQ1M Q1T1 and RND4K Q32T1/Q1T1 writes are pretty much flatlining. Read speeds very consistent throughout the entirety of tests as well, unaffected by what makes SEQ128K and RND4K Q32T16 writes vary so much. Could also be typical behavior.

In the next posts all results obtained, each post for each setting/profile combinations, ordered from the first test with just Windows installed at the top to the second test at 80% capacity filled at the bottom. In the last post a last Verification using the simplified “test it all in three runs” method, and a last few comments for those interested.

Default

Default / Peak

Nvme SSD

nVME / Peak

Real World

First the verification. Meanwhile I stumbled upon the CDM thread, so also adding a Verification using the unified settings there, that I will also post over there.

I’d like to add that I’ve found it pretty interesting to open Task Manager Performance tab and pin up the OS drive there. For one thing, you can see clearly what the drive is aiming at while CDM is running and how consistently it manages to sustain it. These were made while the “Verification 2/3 – Default/Peak Performance” test just above was taking place.

As I’m posting this I’m preparing to go back to Windows 10. I cannot truthfully form an opinion on Optane speeds on Windows 11 because of the registry bypass allowing my PC to make the upgrade. I’ve also briefly installed Windows 11 on my Samsung Evo 970 and it had the same effect: POST sequence is “abnormally” lagging. I didn’t linger much on the 970 EVO Win11 install since I got the Optane the day after and did another clean install straight away. But I must ask myself: what OTHER effects could be happening day to day once the OS is loaded for similar reasons? I’ve not noticed anything out of line there, but perhaps that’s because Optane is indeed very snappy. So I need to go back to my old environment where I’ve got years of experience with the 970 EVO, and see how Optane feels there.

I’ll repost some tests from Win10 but I don’t expect those to be much different. Just while I’m there, you know, I’ll add to the dataset... And by when I’ll post these I should have a final opinion on IF the 1600X brings perceptible OS uplift.

I hope that from here, some of you will post some results from their drives and hardware running it.

 

[mesaprotector]

I hope it's okay to make an account on this forum just to reply (I mean, it's a recent post! ). I bought the same SSD last week and am now running it in my laptop alongside the included Samsung PM9A1, and was curious to find other people's results. My configuration varies a lot from yours - I benchmarked it in Linux with kdiskmark, which is supposed to be a CDM clone, but is not based on the same code.

Didn't take any screenshots but here is the copied results:

KDiskMark (3.1.4): https://github.com/JonMagon/KDiskMark
Flexible I/O Tester (fio-3.36): https://github.com/axboe/fio
--------------------------------------------------------------------------------
* MB/s = 1,000,000 bytes/s [SATA/600 = 600,000,000 bytes/s]
* KB = 1000 bytes, KiB = 1024 bytes

[Read]
Sequential 1 MiB (Q= 8, T= 1): 1737.208 MB/s [ 1696.5 IOPS] < 4681.29 us>
Sequential 1 MiB (Q= 1, T= 1): 1482.394 MB/s [ 1447.7 IOPS] < 688.87 us>
Random 4 KiB (Q= 32, T= 1): 1617.079 MB/s [ 404269.9 IOPS] < 78.30 us>
Random 4 KiB (Q= 1, T= 1): 483.381 MB/s [ 120845.5 IOPS] < 7.98 us>

[Write]
Sequential 1 MiB (Q= 8, T= 1): 1085.037 MB/s [ 1059.6 IOPS] < 7162.34 us>
Sequential 1 MiB (Q= 1, T= 1): 680.941 MB/s [ 665.0 IOPS] < 1014.63 us>
Random 4 KiB (Q= 32, T= 1): 980.418 MB/s [ 245104.6 IOPS] < 128.83 us>
Random 4 KiB (Q= 1, T= 1): 302.238 MB/s [ 75559.6 IOPS] < 12.09 us>

Profile: Default
Test: 1 GiB (x5) [Measure: 5 sec / Interval: 5 sec]
Date: 2024-02-24 19:05:02
OS: arch unknown [linux 6.7.6-arch1-1]

Configuration:
Lenovo LOQ 15IRH8
CPU: Intel i5-13420H
RAM: 2x8 DDR5 5200 MHz (both sticks are 5600 but downclocked by the CPU)
OS: Arch Linux
Filesystem: ext4
relevant BIOS changes: light undervolt on the CPU


The drive was about half full when tested. I wonder about your latency numbers, since Intel doesn't provide latency standards for the P1600X, but for other Optane drives it's usually supposed to be around or less than 10 µs.

My boot times are unchanged (about 12 seconds). It seems like the system is slower to POST but faster in the later steps of the process. And yes, everything feels very snappy, but I wouldn't bet my life it's not the placebo effect making me think that.

 

[QcRMoD]

My guess is that since first your laptop is legit Win11 you won’t feel the lag I do in POST.

My observations are: you run a faster Rnd4K Q1T1 than I do, I run faster sequential than you do. Seems your more recent platform/better hardware optimizes it a good bit. However surprised your Sequentials are relatively slower than mine. I’d have guessed all would scale accordingly.

Cannot inform you further but really glad you subscribed and added to this.

Welcome and thanks!

 

[LabRat 891]

I currently run a P1600X 118GB as boot drive off CPU-connected m.2. I 'split' the drive, and dual-boot WS2019 and Win11IoTEnt off it.

I'll have to re-read through and investigate a few things mentioned...

 

[unwind-protect]

I have one, too. In Linux and FreeBSD is gives random sectors up to 8 times faster than my SN850x. Very happy with it. Will serve to hold my Chrome profiles.

 

[LabRat 891]

Query:
Have you tried RAID0'ing any of these drives? (Off CPU/Mobo; not like ZRAID, etc.)

At least using AMD's NVME RAID on X570, I reliably was experiencing severe performance degradation for each drive I added to the stripped array.
IIRC, 'messing with' the cachestripe size and the cluster size on the formatted array, did effect the scores.
Also IIRC, the 58GB P1600Xs were 'less bad' than the 118GB units (I have at least 2x 58GB P1600Xs, and 3x(RMA on #4) 118GB P1600Xs, for retro experiments and boot drives on Appliances)

I'm curious if that's a symptom of the P1600X's design, or an AMD-specific issue.
(Afterall, they're marketed as single-device Boot and Cache drives)
Last time I messed with that, was a year or 2 ago.
IIRC, my cursory research seemed to point me towards it being firstly an AMD issue, and secondarily Intel doing something to the P1600Xs that advantages only official VROC use.

 

[QcRMoD]

I don’t know anything much personally.

Just published the tests cause I thought there were very little else about them I could find from trusted sources.

 

[LabRat 891]

I don’t know anything much personally.

Just published the tests cause I thought there were very little else about them I could find from trusted sources.

Truth.

The general gamer-enthusiast and even SOHO user kinda ignores Optane's existence.

Those that don't ignore it (and share),
usually get flamed for spending so much $ on such little capacity, and the discussion (often) gets derail'd by those that have no idea of 3DXpoint's and PCM's unique qualities.

 

[QcRMoD]

I… well… not to get flammed in anyway whatsoever… but I think I get some of their idea. I mean everybody does, and that goes the other way around (they do get the general idea with Optane as well), and I don’t think anybody is especially wrong here.

Getting ahead of myself: still have a few things to add to this post before considering I did the best I can with it.

 

[LabRat 891]

I… well… not to get flammed in anyway whatsoever… but I think I get some of their idea. I mean everybody does, and that goes the other way around (they do get the general idea with Optane as well), and I don’t think anybody is especially wrong here.

Getting ahead of myself: still have a few things to add to this post before considering I did the best I can with it.

No worries, and point well-taken.

TBQH, I get 'personally invested' in new technologies, sometimes.
Seeing people far smarter than I write about the very same applications I've day-dreamt of for a given tech, makes me a lil giddy inside.
(Ex: Intel very much wanted 3DXpoint for integration with AI-focused neural-like computing. From what I've gathered from Investment Researchers, there's an IP War going on for AI-focused/specific memory technologies, that the consumer-enthusiast can but see a gleam of)

With Optane in-particular, it has a very intriguing 'story' to it.

1 tid bit:
3D Cross-Point Memory (as Optane is), isn't especially new.
BAE Systems (Yes, the Euro Mil-Ind. Defense Contractor) has been using "Ovonic/Ovonyx" PCM in certain aerospace systems, for years. That "Ovonyx PCM" *is* derived from the exact same patents and technology that Intel(and Micron) more/less pulled out from under the inventor (as he was passing away, no less).

To me, even merely owning Optane is like having a bit of a Cyberpunk International Corpo-Drama, while being legitimately a 'Mil-Spec' Aerospace Technology.
-and, that's just Cool AF

 

[mesaprotector]

I bought one for the same reasons - incredibly cool technology that might not be available much longer, and a spare M.2 slot on my laptop I wasn't planning on using anyway. I also have heard putting Linux/W11 on separate drives is a good idea, though moving my install over was incredibly stressful - it worked out! I figure with the crazy 1.26PB advertised endurance, I might even be able to use this as a boot drive on my next machine years down the road.

Wish Intel would've made a follow-up to the 905p; SSD technology is one of the few areas of hardware that seem to be going backwards. Solidigm (Hynix's new brand) is coming out soon with an SLC U.2 drive - everything else on the market is TLC NAND or worse. (Though I will say that current NAND is better for stuff like a video collection, where sequential writes matter a bit.)

 

[Selaya]

Query:
Have you tried RAID0'ing any of these drives? [ ... ]

i can already tell you whats going to happen: random performance (the metric you'd care for) will actually get worse. don't do raid0 w/ ssds. remember raid0 = 0brain

 

[QcRMoD]

UPDATE

A lot of stuff went down since the OP, so I’ll try and be succinct where I can.

Going back to Win10 for about a week just pretty much verified my assumptions about the Optane – I’ll save us a lot of pictures here and just present the compilation I made of CDM testing there – one run with the Windows clean install, one run artifcially filled to 80%.

Meanwhile my best man running an old 3570 non-K from an ancient HTPC of mine decided finally to upgrade lately. I’m introducing results from the SN850X I got him running from an i5 12600K, Z790 Riptide and 32GB 5200Mhz RAM.

At the point I tested the SN850X (above the best run out of two), Win11 was updated and de-cluttered and the drive overprovisioned and there were some programs and bench installed. I think it fairly represents nVME 4.0 speeds in the broad and will be used as such.

Meanwhile still, I reinstalled Win 11 on my PC for good on the Optane, so here are final testings but do remark that in the final tests of the Optane my basic environment was fully installed with no artificial fill-up there, ending around 41% capacity:

This in turn yields us a final grid of the Optane with the full dataset including those above:

Then I got the idea of a wider range comparison: plugged my 970 Evo 250GB 3.0x4 in there (my OS drive until the Optane, which I intend to use in my HTPC now) and overprovisionned and artificially filled it up trying to reciprocate the SN850X state when tested – it all went in hand as best as I could improvise it. It should serve well for nVME 3.0 speeds in the broad.

Then tested my two old Sata drives I use for light data (the heavy stuff is on the media server) and some games. Coincidentally – none by design but by mere chance – they were each filled around both states I tested the Optane in my first post and going back to Windows 10: the Crucial was at 31% and the Samsung just over 80%. As the slightly faster of the two, the Samsung was elected to represent Sata SSD in the all-comp, but it could have been the other way around without changing anything.

Did not take captures for these: I basically filled up the all-drives comp as the tests went to save time.

I want to point out again, as I warned in the OP, that the Optane results tended to vary more wildly in Sequential Writes. I couldn’t say much about it then, but since and with all the tests carried I’ ve found all three nVME tended to similarly behave in Writes more than Reads or Mix (although there is some of that in some cases too) specifically around the same tests: SEQ128K Q32T1, SEQ1M Q1T1.

The same can be said of RND4K Q32T16, where all nVME exibited a lot of variance in Writes. Didn’t seem to matter if the drive was OS or just a Data drive but happens anyhow. Where Optane is very very stable, almost flatlining, in RND4K Q32T1 and Q1T1, I remarked that at least the Q32T1 has “regular” nVME vary a good bit as well. The Sata SSDs really are the least touched by all of these variances.

Looking at the Performance tab in task manager, you can actually see it happening and it looks like throttling – not from temperature though I believe. All drives tested here were a pretty cool 50*C at worst running CDM (HWinfo). You can see it at the start of the test pushing to it’s highest range, but it doesn’t manage to settle there, goes all over the place to extreme lows yielding a poorer result. Then you redo the test and there it’s pushing to its max and managing to keep it thereabout. I have no explanation to give but it’s there, hence why I ran all later drives tested at least twice and used the best.

The clarity brought by all the testing, and especially testing Optane so frequently, makes me want to add something I didn’t say in the OP because I thought I’d look stupid, yet the happenstance was so clear and localized and it shows so strongly in the first results I published that I cannot do otherwise now. The Optane was brand new when I did the first Win 11 install as is. The first batch of tests I did with just the clean install had lowish Sequential/RND4K Q32T16 Writes. Then when I filled it up to 80%, I started seeing a lot more of it going at and over 1000 MB/s, but the drive still performed below that in many occurences and repeats. Sounds ignorant but it seemed to me that it sort of needed to “open up”.

Then I went back to Windows 10, which meant erasing the whole drive, and finally started to really get consistent results. The typical variance was there but I never encountered 700-800 MB/s runs of Writes there. Went back to Win 11 erasing the drive again, get the same consistency now. Typical variance, no 700-800 MB/s runs. I’d have liked to test the brand new SN850X further as well, see if I could observe the same phenomenon, but I had a couple hours of benching/stressing ahead of me and wanted it done.

So as much as I’m happy with the final Optane compilation grid for being extensive and that it at least got to a point where that initial “sluggishness” with Writes was averaged by subsequent runs to almost reflect the real performance once it started to behave consistently, it does not reflect at all the Win 10 compilation grid or the final tests I did on the second Win11 install. So for the all-comp grid below, I averaged the Win 10 grid and those last three tests – a dataset of 3 runs at three different levels of fill for each test. All other drives were tested at least twice and I retained the best scores for these.


Closing thoughts on the Optane 1600X

I’ll cut short the BS about what my experience is: yes it makes for a real snappy OS, and yes in reinstalling Windows three times there and also installing it on the SN850X, it’s clear the lead time for Windows to set things up before actually getting to the Desktop, the speed at which it dispatches with installing all the initial updates, and with uninstalling all the crap, and then some, the Optane is faster.

Knowing that, but also knowing that a good regular nVME is no slur and is still fast in a make-you-happy way, it’s clear the ratio of such performance vs. overall value is what didn’t work so well for Optane for consumer market adoption. A good nVME 4.0 costs little more than the 1600X 118GB to give you 1TB. If anything like a tight budget or just the wisest investment for use is at stakes, there’s a fair chance one would get a 1TB nVME 4.0 unless there was a critical productivity/OS bound benefit to the Optane.

In the end, when Windows is installed and all our favorite stuff as well, the OS is most often used for most people as a gateway into stuff which performance is more strictly bound by CPU/RAM/GPU or even Network, sound cards, monitor, and what have you. Like, nothing any nVME that makes sense cannot do well of whenever it’s needed. Just like the first consumer-oriented Optane were cache drives to the OS drive, the optic of a consumer Optane SSD was never capacity. Not geared towards ultimate Sequential speeds because really if it would push like 4 GB/s, you’d fill it up in 10 secs – and then WHAT? So what it does is level modest speeds at both - Reads fast enough to basically read the whole drive within 1-2 minutes, Writes fast enough to fill it all within 2-3 minutes. Pushing 7000 MB/s over 1TB, you'd read it entirely at best within 2 minutes. See where sometimes, it's good to have blazing fast Sequentials, but that there is still a limit - and a design - for them to be useless?

So neither capacity is sufficient nor top speeds warranted on bigger data blocks because it’s not the point. By design, the Optane 1600X isn’t meant to be anything more than strictly OS and everyday/critical apps. I mean, even some single games nowadays could not fit on an empty 118GB. But beyond understanding that, it’s not EVEN meant to be strictly performant on all-Reads or all-Writes stuff. Not only in the lowest RND4K Q1T1 where Optane is fastest, but all through testing, the Optane is obviously geared to be most efficient with Mixed workloads.

It becomes apparent in the grid indeed that most other drives tend to level down to their worst speed of either Reads or Writes in most Mixed workload. The Optane always levels up to the best it can do there, so much so that it starts to realistically equalize Mixed performances from ballpark nVME 3.0 RND4K Q32T16, to those from ballpark nVME 4.0 in Q32T1 – and THEN it basically KILLS every regular nVME out there in Mixed Q1T1 as well as it kills them in Reads only. If you look at all the other drives, somewhere between SEQ1M Q1T1 and RND4K Q32T1, they even start to get lower Mixed speeds than either their pure Reads or Writes can get at. The Optane is also the longest to keep with a much more effective percentage of its Peak speeds as the data blocks get smaller/QT tighter.

That is what Optane does best, that is its value and its performance. If anyone wants to milk EACH step of a PC integration for the best performance, the Optane is the necessary OS drive still available today – but by design it is also an extra expense because anyone will want the best speeds they can get on at least ONE much bigger capacity drive. It’s essentially what it was at the start: a cache drive, but now with an accomodating enough capacity for the full OS and a lot of heavy footprint apps.

I’m glad then that I’ve chosen to run all CDM tests with a Mixed workload as well. It’s seldom seen out there, and where CDM results can tell anything, further serves to hide where even a 1600X will shine. Then, beyond synthetic benches even if carried to better underline its strengths, there are other advantages: longevity, learning from the workloads it most often gets, basically impervious to performance degradation until writing it into it’s last couple GBs space left I guess.

Thanks to @dgianstefani for basically convincing me I should absolutely try one. I’m not so sure I’m the best candidate for it, but it was a fun ride and a very enlightening process to test and compare it. I also happen to love it. Another thing the 1600x shines at is kicking this older platfrom of mine closer to par with my best friend’s new rig with the “faster” nVME and truly much faster components. My only regret is not having a decent opportunity to experience the 1600X within such a much more modern platform.

 

[LabRat 891]

I'm staggered, in a good way. Thank you so much for all this testing!
I'd been hoping to see some in-depth analysis on the P1600X; they're both surprisingly-available for an officially 'dead' product.

The clarity brought by all the testing, and especially testing Optane so frequently, makes me want to add something I didn’t say in the OP because I thought I’d look stupid, yet the happenstance was so clear and localized and it shows so strongly in the first results I published that I cannot do otherwise now. The Optane was brand new when I did the first Win 11 install as is. The first batch of tests I did with just the clean install had lowish Sequential/RND4K Q32T16 Writes. Then when I filled it up to 80%, I started seeing a lot more of it going at and over 1000 MB/s, but the drive still performed below that in many occurences and repeats. Sounds ignorant but it seemed to me that it sort of needed to “open up”.

I'd noticed this to a degree, also.
Considering that Optane-3DXPoint is more like DVD-RAM than it is NAND, I could actually see some kind of 'exercising' having an effect.
(what driver used will impact perf. IIRC, I manually extracted and installed the Intel driver for my P1600X.)

Also, thermals:
IIRC, Optane cells get slower the cooler they are below like 40-50c. But, the Power ICs on the Optane drives get searingly hot, and the nearby controller does need to stay nominally cool.

That is what Optane does best, that is its value and its performance. If anyone wants to milk EACH step of a PC integration for the best performance, the Optane is the necessary OS drive still available today – but by design it is also an extra expense because anyone will want the best speeds they can get on at least ONE much bigger capacity drive. It’s essentially what it was at the start: a cache drive, but now with an accomodating enough capacity for the full OS and a lot of heavy footprint apps.

Your efforts and testing confirms my feelings and inklings about the P1600X:
It may not look like a good performer at first glance but, in actuality is one of the easiest upgrades to 'responsiveness' you can give your computer.

Largely in thanks to you,
I will be keeping my 118GB P1600X as boot drive in my CPU-connected slot, even though I could use more storage space and will soon be losing access to an X1 slot (2.5+ slot GPU on-order)



The only thing all this testing didn't cover (which, for publishing's sake, is good) is how formatted Cluster Size can effect performance on Optane.
Personally, I've noticed larger clusters, slowed down the P1600X 118GB.

 

[QcRMoD]

I thank you very much for your interest and participation.

The only thing all this testing didn't cover (which, for publishing's sake, is good) is how formatted Cluster Size can effect performance on Optane.
Personally, I've noticed larger clusters, slowed down the P1600X 118GB.

Seeing how it CAN deal with as little 512B segments, this is probably expected. ATTO SSD benchmark can be interesting in that regard.

But asking synth benchmarks to tell the whole story is probably useless. Also I do lack comparisons with ATTO of the SN850X or the 970 Evo.

I'd like to see one proper review/expansive sets of tests carried on it. But I guess what I did is the best we'll get out of it until someone with the tools and knowledge to really say/show some things about it will arise.

 

[dgianstefani]

I did say I'll bench my drive on my more modern system as a point of comparison, but I've been too busy recently.

Something to bear in mind, that isn't immediately obvious from the results in this thread, is the sustained consistency you get with Optane. This might be a little hard to see with a 118 GB drive, but it's obvious in the P5800X benchmarks. What happens with short bursty benchmarks like ATTO and Crystaldiskmark, is that NAND flash SSDs can just use their pseudoSLC cache. Here's what happens when that runs out.

Intel Optane SSD DC P5800X Review: The Fastest SSD Ever Made

The fastest just got faster

www.tomshardware.com

The flat lines are Optane.

Even without a bursty workload, if you simply fill a NAND flash drive up to 80-90%, it won't be able to use that pseudoSLC cache, since it takes 3-4x the capacity to operate.

So, the 905p is a good all around upgrade from nand flash, the P5800x is a halo product, and the P1600X is entry level, faster than the 905p in everything except sequential, but limited in capacity.

 

[mesaprotector]

KDiskMark actually claims I'm slower on mixed workloads, but I almost wonder if KDiskMark is calculating it incorrectly, because my "mix" latency is very good, but all the MB/s numbers are right about half what they should be. I could definitely see someone making that programming error. Until proven otherwise I'm going to choose to believe that. Unfortunately no CDM on Linux, and I don't feel like adding a Windows partition just to benchmark properly.

LabRat 891 said:

Also, thermals:
IIRC, Optane cells get slower the cooler they are below like 40-50c. But, the Power ICs on the Optane drives get searingly hot, and the nearby controller does need to stay nominally cool.

Just to add a data point, my drive is very consistently around 36-38C during idle/web browsing, about 8 degrees warmer than the Samsung. That part of the keyboard does feel rather warm, but I've never seen the drive above 52. I don't game on Linux so I can't really test that aspect of it. My laptop's idle power draw is up by 1.1W (extremely non-scientific test; I just closed everything and left it for 5 minutes with the screen on minimum).

One weird aspect of using a drive that's extremely fast at random ops but kinda slow at sequentials is that my time to resume from hibernation is now slower than my time to boot up from zero. I only use hibernation as a failsafe if I forget to plug my laptop in, so I don't mind.

EDIT: Tested my Samsung as well, and it really does look like a KDiskMark bug.

 

[QcRMoD]

KDiskMark actually claims I'm slower on mixed workloads, but I almost wonder if KDiskMark is calculating it incorrectly, because my "mix" latency is very good, but all the MB/s numbers are right about half what they should be. I could definitely see someone making that programming error. Until proven otherwise I'm going to choose to believe that. Unfortunately no CDM on Linux, and I don't feel like adding a Windows partition just to benchmark properly.

View attachment 337507

You're SEQ Q1T1 is also far off grid - way slower. But the RND4K Q1T1 R or W are exceptional. By latency the RND4K Mix iOPS/MB/s however is indeed not expected. Looks like it should have scaled toward 380 MB/s or something.

I therefore have a hard time reconciling your results in a comparable way with mine.

Unless you did only one run. Do a few and preserve the best run to show.

 

[mesaprotector]

I did do a couple runs, with different presets, and it seems consistent. The first post I made, I got 681 MB/s on SEQ1M Q1T1; this one was 686. I don't know why that test specifically is so slow. Most other sequential tests, I'm consistently only about 2-3% slower than you are. I just did a new test for SEQ1M Q8T1 and SEQ128K Q32T1, and I get 1740/1743 MB/s read, 1085/1084 MB/s write. Actually a reasonably good match for Intel's own specs:

As to the mixed results, I'm guessing KDM is accidentally only counting reads, or only writes, so the real numbers should be exactly double the ones it gave:

SEQ1M Q1T1 rw: 976 MB/s
RND4K Q1T1 rw: 396 MB/s
RND4K IOPS rw: ~99,000

I also get very slow results on RND4K Q32T16... maybe because my processor only runs on 12 threads?

 

[QcRMoD]

I've personally not seen Optane having any much variance in RND4K within something like 10 MB/s and congruent latency/IOPS. In pure Sequential Writes I'm sure you should get better. In Reads expecting 1500+ but you're about up speed there. Seq Mix with proper Writes should land about 1200 MB/s. That on Seq1M Q1T1.

I've not even tapped into "absolute" Optane SEQ as claimed by Intel. I'd be averagely 70 MB/s slower, also applicable roughly to my RND4K Q1T1.

On one side your RND4K R or W makes me envious, as do your latencies OVERALL. For the rest it's not awful but from your RND4K I'd expect you could tap max SEQ and more importantly way faster Mixed speeds.

My system as per account info is 4/8 from 2017 with expected 3200MHz range DDR4 2x8GB.

Another explanation could be you didn't read my entire thread (quite sane behavior) and missed the bout where I said I had shitty Writes/Mixed and even relatively slower overall perf from the brand new unallocated Optane with a direct Win 11 install. Things got better when I artificially filled the Optane to 80% on top of the OS. Then even better when from there I fully erased the drive and reinstalled both Win 10 and later Win 11 after a second erase. If the laptop is new and original OEM install barely over 50% filled then I can see a perfect case I guess of your results.

 

[mesaprotector]

I've personally not seen Optane having any much variance in RND4K within something like 10 MB/s and congruent latency/IOPS. In pure Sequential Writes I'm sure you should get better. In Reads expecting 1500+ but you're about up speed there. Seq Mix with proper Writes should land about 1200 MB/s.

I've not even tapped into "absolute" Optane SEQ as claimed by Intel. I'd be averagely 70 MB/s slower, also applicable roughly to my RND4K Q1T1.

On one side your RND4K R or W makes me envious, as do your latencies OVERALL. For the rest it's not awful but from your RND4K I'd expect you could tap max SEQ and more importantly way faster Mixed speeds.

My system as per account info is 4/8 from 2017 with expected 3200MHz range DDR4 2x8GB.

I double-checked all the results I have and I think I found the culprit - my OS! I can't benchmark the Optane on Windows 11, but I already did for the Samsung, and it's muuuuuch faster on sequential Q1T1 writes there. Like 40% faster. This doesn't happen for reads, just writes. I don't know what exactly about Linux is causing that (the filesystem? some missing driver?) but I can deal with it later.

I can definitely try filling the drive to 80% (reinstall is not happening)... but I feel if I'm getting the same problem with the 6-month-old Samsung 1TB, that probably isn't gonna help.

 

[unwind-protect]

I don't know what exactly about Linux is causing that (the filesystem? some missing driver?) but I can deal with it later.

What filesystem do you have?

Might be worth trying a couple different ones.

Driver issues are unlikely with NVMe.

 

[QcRMoD]

I double-checked all the results I have and I think I found the culprit - my OS! I can't benchmark the Optane on Windows 11, but I already did for the Samsung, and it's muuuuuch faster on sequential Q1T1 writes there. Like 40% faster. This doesn't happen for reads, just writes. I don't know what exactly about Linux is causing that (the filesystem? some missing driver?) but I can deal with it later.

I can definitely try filling the drive to 80% (reinstall is not happening)... but I feel if I'm getting the same problem with the 6-month-old Samsung 1TB, that probably isn't gonna help.

You don’t have to reinstall I would surmise.

From experience and thinking about it I think one could:

shrink the drive to any extent permited
fill whatever left
Test just to see
Clean the extra partition
Extend the OS partition
Test for sure consequently


As is then I'd like to see scores from:

SEQ128K Q32T1 + RND4K Q32T16 + iOPS + latency
SEQ1M Q8T1 + RND4K Q32T1 + iOPS + latency

Just to get an idea for myself of what I somehow also encountered - as it translates with your previously posted scores from KDisk.

 

[mesaprotector]

What filesystem do you have?

Might be worth trying a couple different ones.

Driver issues are unlikely with NVMe.

Plain old ext4. No encryption.

I'd rather not run out my drive's endurance testing it (ha, who am I kidding, it's an Optane, it'll never run out of endurance ) but this got me curious. I'll include all of my sleuthing in a spoiler box below, because this is super niche stuff and I don't want to wear out my welcome right after arriving on the forum.

Spoiler

It turns out it's fairly easy to find people online who have similar results to me on Linux - normal performance at Q8T1, reeeeeaaaallly bad specifically for Q1T1 write.

Specifically I found a thread where people are benchmarking the SSDs on their Framework laptops, and a number of people are similar. But not everyone! And it doesn't seem dependent on distro either. Some random Korean page had the same thing happen, and a random Linux tutorial webpage too.

That last one, though, really looks weird, because they did something the other pages didn't - included the text file of the results instead of just the image.

Notice anything? It's the same situation as I noticed above with my seemingly slow mixed speeds - the latency numbers don't match up to everything else! Specifically, the sequential 1M Q1T1 read latency is almost identical to the sequential 1M Q1T1 write latency - but the speed for writes is half! I could be wrong to think of this as weird, but at the very least, this doesn't happen in CrystalDiskMark (following are the results for my Samsung drive there):

The latency numbers CDM provides match up to the rest of the data. The SEQ1M Q1T1 results are close to each other, and so are the latencies. The SEQ1M Q8T1 write has a similar transfer speed, but being at a queue depth of 8 the latency is roughly 8x that of the Q1T1 numbers. All the rest of the table works out the same way.

Not so for KDM. Based on my latency numbers, my SEQ1M Q1T1 write "should" be 995 MB/s, and mixed "should" be 1188 MB/s. But they aren't. I'm not knowledgeable enough to know if this is a KDiskMark bug, or "real" thanks to some weird timeout some Linux utility is throwing into file operations - either way it happening only at Q1T1 is very weird, but I'm not especially concerned, since I didn't buy this drive for its sequential write speed.

 

[mesaprotector]

I doubt anyone will ever find this thread in Google search results, but a tl;dr for the above: if you're benchmarking your drive using fio (which KDiskMark uses as a backend) and want comparable results to CrystalDiskMark, make sure you have the option --scramble-buffers=0 and don't have the option --refill-buffers. The short explanation is that KDiskMark, by default, generates new random data for each block it writes, so in addition to benchmarking your drive, you're also benchmarking your CPU's ability to generate random numbers which... isn't really what you want.

A quick re-test after doing the above gives me a couple more reasonable results:

Sequential 1MiB Q1T1 write: 1017 MB/s
Sequential 1MiB Q1T1 mixed: 1230 MB/s
Random 4KiB Q1T1 write: 322 MB/s
Random 4KiB Q1T1 mixed: 388 MB/s

This only affects write workloads, so I didn't bother redoing the read tests.