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

[Fokker]

It's funny I ran across this today. I just ordered a 58GB one to use as drive for an RPi5 project.

 

[QcRMoD]

I like the… lithography but why no picture for these results where not shy of them before?

Asked in good faith. The info is interesting. The explaining seems founded. I’m too lazy getting there is all - until I see validation.

Which with my uneducated means I tried to provide indeed.

 

[mesaprotector]

I like the… lithography but why no picture for these resulsts where not shy of them before?

Asked in good faith. The info is interesting. The explaining seems founded. I’m too lazy getting there is all - until I see validation.

Which with my uneducated means I tried to provide indeed.

Sorry about that, I had to run the tests in the command line so no pretty pictures - though I'm happy to validate by screenshotting my terminal, if that works for you:

 

[QcRMoD]

Sorry about that, I had to run the tests in the command line so no pretty pictures - though I'm happy to validate by screenshotting my terminal, if that works for you:
View attachment 337988

Lol prettiness holds little interest to me although I do like to make an effort for clarity… still your point is FELT here and due apologies…

…now I like this. Will look into it and may provide further pretty pictures if I can add something of interest.

Thanks for further validation. I’ve really appreciated your precise take here.

 

[QcRMoD]

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.

I’ve gone a bit into it now – trying to understand what your additional infos meant exactly.

One interesting answer on StockExchange to a member asking for a comparable test to CDM on Linux:

Particularly remarkable is it has been posted 10 years ago and the given answer, after searching (quickly but using a few different queries and reading about a dozen output from these), seemed to outline a basic notion about CDM much better than other sources...

The same notion, to some extent, correlates to how Nutanix explain why CDM “0Fill” option should not be used “to be fair” in testing:

CDM Info is quite succint on the matter, but what they do have to say would work well to color Nutanix input AND again underline how simple and perfect AND nuanced the answer given on StackExchange:

As per how you tried to sum up your explanation, I’m not entirely sure you got it right. However, I’m not entirely sure I get the gist of it right neither, so just to open up discussion.

The “Default” dataset for CDM is Random indeed. As per the StackExchange answer, it does seem fio, which Kdisk uses as backend, would be the closest to reciprocate CDM. I didn’t find a comparable explanation to yours with more details so that I could try to understand it better.

But from my general understanding of these above and other stuff:

-As per CDM, some SSDs will scale better with a specific dataset (to 0 Fill or to generate Random data)... which truncated form I would take it to mean that some SSDs will NOT so much.

-As per Nutanix, the Default Random in CDM should be prefered to get a fair comparison throughout.

-As per that guy on StackExchange, fio would seem to be an option that would generate similar workload, BUT that a difference is in how CDM can’t do raw I/O but creates a filesystem – which gives its scores that particular filesystem overhead which might make comparing different results difficult. At this point I am not particularly sure if fio uses a filesystem to be comparable to CDM, or does raw I/O in some specific way to be “comparable” to CDM, or if your additional setup is a way to get it even closer.

But it does need to be further defined – a task I’m not saying YOU should have to do, but one anyone willing to participate might try to clarify.

There are also suggestions to benchmarks that work both on Windows and Linux. Then again I guess it’d be rather difficult to make sure how reciprocating they can be within two different ecosystems.

Enough said... I decided to test the Optane with the 0Fill Dataset from CDM just to see. For those with a sharp eye, well then yes the drive got filled a bit further naturally from setting up my exacting ecosystem.

Also, I am currently running 32GB of RAM instead of 16GB as part of a test as well as a partial move preparing for a new core to my PC which is already ordered and should get to me early next week. So THEN I will be able to post Optane results from yet another Win11 install AND an up to date PC.

Also, for the two first runs I had DisplayCal and my monitor profile loaded and was listening music on Spotify while they ran. For the Verification run, I disabled the calibration, closed both apps, rebooted, and retested “bare” as I did for all of the earlier tests. As anyone will be able to figure out where the Optane is concerned, neither the 0Fill dataset or the fact that Spotify was running or that I have double the amount of RAM or a monitor calibration profile and app running background changed ANYTHING much outside of regular variances. In fact as far as DisplayCal is concerned, it never impacted anything since 7 years I’ve been using it – even gaming.

...

Aaaaand compiling the 0Fill runs...

As part of further tests as well (not in this post) and while still running the same 7700K / Maximus IX Code core system, I will “0Fill test” the Samsung 970 EVO and 850 EVO just to see if any of them DOES scale one way or another with that dataset. When I’ll be running the newer platfrom, I’ll retest all three drives under both dataset. And then I’ll take a leave of CDM for another decade I suspect.

 

[mesaprotector]

I’ve gone a bit into it now – trying to understand what your additional infos meant exactly.

...

There are also suggestions to benchmarks that work both on Windows and Linux.

I appreciate your perspective! I'll respond to only a couple points - fio does actually work on Windows, though hardly anyone uses it because CDM is pretty standard (along with ATTO and a few others). I tried the same fio commands KDiskMark uses in windows, and got the exact same problem - unusually slow SEQ1M Q1T1 write (relative to CDM), everything else normal. Though the difference was smaller than on Linux - about 15% too slow instead of 40%.

Fio too has an option to write zeroes, but I set that to false in all of the tests I posted. I did try it once - and for me, at least, the results were identical to writing random data.

I'm done with SSD benchmarking for a while, personally (and probably with CPU/GPU benchmarking as well, unless someone beats my Time Spy score...). I had a few bugs come up this week, unrelated to the testing but they made me realize I should just appreciate my laptop working normally. I'd love to see you and more people post any info they come up with, though.

 

[QcRMoD]

0Fill – Samsung 970 EVO (PCIe 3.0x4) & 850 EVO (Sata)

*Sorry for an error on the original all ssds comp grid, where the Samsung sata was labelled "950 EVO" which of course should have been "850 EVO"... since a 950 EVO doesn't exist and its PRO variation is a nVME. To my pitiful defense there was a LOT of data and I really think Samsung are way overdue to change their monikers for their nVME and especially stop reusing the same numbers into a different series 2 years after...

Both drives overprovisioned the same, and filled with the same exact data – which is the original data that was already on the 850 EVO. Therefore in earlier runs I posted the Sata was 84% as it is here, but the nVME was overprovisioned and filled to better match with the SN850X 1TB I had tested.

Without further ado... the (corrected) all ssds comp grid, so that we have our reference at end. To clarify those were run with the Default (Random) Data Set.

Helpful here as I don't intend to make another one of these for the 0Fill especially since I cannot retest the SN850X, and especially since a new comp grid would only be useful if I expanded it to include the above and the 0Fill side by side. That's where I draw the line. Instead I'll use pictured of the CDM results in 0Fill and a reminder of what we "expected" based on the Default Data Set.

970 EVO 250GB 3.0x4

nVME Peak - expecting 3564/1507/1601 in SEQ128K and 1035/1531/1108 in RND4K

Default Peak - expecting 3567/1537/1819 in SEQ1M and 716/488/576 in RND4K

Real World Performance - expecting 2383/1538/1425 in SEQ1M and 71/210/78 in RND4K

850 EVO 250GB Sata

nVME Peak - expecting 555/520/480 in SEQ128K and 401/326/378 in RND4K

Default Peak - expecting 553/532/502 in SEQ1M and 401/362/378 in RND4K

Real World Performance - expecting 511/496/446 in SEQ1M and 37/149/60 in RND4K

Hmmmm... okay so we have outliers, but they go both ways - slower Default Data Set OR slower 0Fill Data Set. And they sort of disseminate at all places. Instead of considering it an effect of 0Fill, some elements make me suspicious. Therefore I must validate - in as trying to reciprocate results from both Data Set - for a few scores because they rather are illuminated for being out of place.

970 EVO outliers:

-in nVME Peak the 0Fill is much slower in RND4K Mix. Looking at the similar Reads and Writes with both Data Sets, the gap will surely close in validation.
-in Default Peak the Default Data Set was quite slower on all RND4K. That would SEEM to go in phase with what is "expected" of a 0Fill run but I'm very curious to validate.
-in Real World the Default Data Set was much slower on both Reads and Mix.. More or less goes with "expected" of a 0Fill... Writes would usually be the uplift instead of Reads. Curious to validate as well.

Let's not forget also that the 970 EVO on the first batch of test was filled around 6% instead of 84%. However my instinct is all the validations might close most of those gaps until we cannot speak of an outlier anymore, despite the huge difference in drive fill - it's too all over the place to sound like some form of scaling gained by not using random data sets.

850 EVO outliers:

-in nVME Peak the Default Data Set was quite slower in Writes. Looking at the similar Reads and Mix results with both Data Sets, the gap will most surely close in validation.
-in Real World the Default Data Set was quite slower in Reads. Again, looking at the similar Writes and Mix results with both, the validation shall prove it an outlier as well.

Going to validate these as I write this.

EDIT:

850 EVO outliers RECIPROCATED on a first run...

All concerned RND4K tests showed as to outline the consistency of ALL results there.

EDIT:

970 EVO plays it harder...

I could reciprocate easily all Default Peak RND4K outliers on the first run.

Not at all after two validation runs regarding the nVME 0Fill RND4K Mix scores... Still as low as posted in the first run.

And that is really weird. I'll try to validate the other way around with running a Default Data Set... cause I think the only sensible explanation is that I made an error when I filled the all SSDs comp grid.

Getting somewhat weirder, I more or less reciprocated the outliers for the Default Data Set Real World SEQ on first run... in that I got EVEN FASTER score than the 0Fill run by the same margin it was slower in the compilation grid... but still the same relatively slower Mix score.

So going full validation mode now on the 970 running nVME and Real World for both data sets a few times and see what I get.

FINAL EDIT

The ONLY explanation for the nVME RND4K Q32T16 Mix HUGE gap was a HUGE error when filling my All SSDs compilation grid of earlier.

Otherwise, EVERYTHING that was problematic between 0Fill and Default Data Set is most strictly flatlining.

Real World SEQ tend to vary a little bit especially towards the Mix, and yeah at times you might get a super boost in Reads like I proved with the first attempt at validation, but for this i7 7700K and yadda yadda core system, it seems obvious the SEQ1M Q1T1 Read target is 2500+ MB/s for a Write target of 1535+ MB/s and a Mix target of 1400+ MB/s. Writes very very consistent, but Read and Mix do get some wider variance.




0Fill nVME Peak vs Default Data Set nVME Peak:

0Fill Real World vs. Default Data Set Real World:

 

[QcRMoD]

“0Fill Report” + Final thoughts from testing it

Well yeah... indeed this report will have nothing really to do with 0Fill or Default (Random) Data Set because, not without some extra effort and realizing I’ve made an error in the earlier all SSDs Comp grid, I’ve also just shown three different drives of different technologies and speeds (including the Optane) where there is negligible differences between running CDM in both Data Set (especially with higher speeds disseminated in both)– once you do have enough runs done to confirm the best speeds the drive can run at and can repeat consistently. Negligible being into decimals at best to representing 1-5% of total effective speed at worst, which is unavoidable variance. Even Optane is not impervious to those as I extensively shown, yet it becomes more of a 1-2% variance kind in its case.

One quirk that I have now observed a dozen of times with the 970 EVO appears readily when running a CDM Peak or Real World test using the “ALL” button. Examples below captured while validating the Samsung 970 EVO results in nVME Peak and Real World profile and in both Data Set – so I don’t think that quirk is dependent of any settings/profile BUT I also do not know if it’s drive dependent or system dependent or something that would affect most nVME Nand in most systems – or something with CDM.

What I believe IS the cause should appear more clearly with the captures. To the left, Performance tab in Windows Manager for the drive. Upper right the CDM test actually running. Lower right is the capture of my original test to exemplify what I was looking to reciprocate or find that I could not.


Validation of nVME Peak in 0Fill...

RND4K Q32T16 Reads... Houston, we have a problem. Otherwise all other tests seem rather congruent.

A second run yields in line result.

No need to dwell on trying to have the absolute HIGHEST CDM scores for EACH result, because variances will appear where they will appear – eg. 1426 or 1462 MB/s, I don’t really care – it’s only about 3% variance. Performance manager would tend to show while that test is running that the drive tries to target 1.4 to 1.5 GB/s and manages to maintain it in long enough bursts that the score reflects it.

What you WOULD observe when things go wrong is that the drive cannot seem to maintain it’s target more than “in passing” – and that it will go through the lowest speeds and many intermediate speeds most of the run. You’d still be able most of the times to identify the speed it tries to target because it’s the highest you’ll see in passing with some redundancy. But it can’t stay put there even for a full second.

And in this case that is not thermal throttling, proof being it settled to optimal speeds if I rerun the outlying test right after the first run is completed.

So that in this other similar case...

Validation of nVME Peak in Default (Random) Data Set...

Same problem happening.

Yet everything else congruent in that run

I don't even need to wait for CDM to start displaying results after iteration 1/5... Performance Manager already tells me the drive is pushing to its target and keeping at it consistently enough...

Then most usually, no worries for the rest of that rerun, scores should align neatly 99.9% of the times. When using Peak or Real World profiles and a rerun in RND4K where most quirks occur, it’s that more obvious because not only speeds are within target, but also because iOPS and latency will be quite strictly congruent as well.



Just another one happenstance, just for fun... no comments.

Validation of Real World Perfomance in Default (Random) Data Set

That quirk made me think... because it only affects the Read iteration of the RND4K test, then usually goes to produce consistent to target Writes and Mix. So it seems to be the “switching” from Sequential testing to Random testing that produces it, and by the end of the Read run and waiting time it seems to finally settle optimally for the rest of the runs. Again, it could be a quirk with CDM but then Optane is not affected by it, which tends to disqualify the hypothesis.

Once I found my bearings in the OP, my focus was to underline how Optane is, at ALL levels of testing, optimized to milk the most of a Mix workload and to be consistent at it. Which in turn, additionally to small data blocks speeds, is what sets Optane apart as a performance for an OS drive.

Coming back to our very obvious Nand quirk now as it shows in articificial CDM testing, let’s extrapolate the concept of Mix to the concept of an OS. Where indeed, you’ll not only get mixed R/W workloads, but you’ll also often get mixed Sequential AND Random workloads.

On the broader scale of EVERYTHING iOPS your OS drive needs to do at all times, Nand SSDs seem to be prompt from their own structure to have difficulties not only maintaining performance at any gear in any Mixed workloads, but also SWITCHING gears between a Sequential workload and a Random workload. Worst even, it seems that once they START on a task with terrible performance, they do not recuperate until that task is done.

I mean I cannot readily test 10 different nVME drives and prove anything discussed above, but something of interest I wanted to share. BTW, if you run those tests “backwards” (not using “ALL” and beginning with RND4K run to immediately run the Sequential test as soon as RND4K is done), there is no similar hitch to the Sequential Read speeds. Not entirely surprising since Sequential is really where nVME shines best, but I must underline there’s no quirk there, perhaps disproving my theory.

Finally the "corrected two times" all-SSDs comp grid.

 

[QcRMoD]

As a final to this thread, I'm looking at both the Optane 1600X and the 970 EVO when tested on my new core system.

So for all the other posts I was testing with a 7700K - M9C - 16GB 3200MHz C16/32GB 3000MHz C15 DDR4 (where the extra RAM borrowed from my now former server didn't change anything). All of these were ran obviously with the PCIe 3.0x4 capacity from that Z290 motherboard.

The results below were obtained on:

i5 14600KF (no OC - XMP Profile loaded)
Z790 PRO RS
32GB 6000MHz C30

Where the Optane was concerned, everything straightforward as usual, except of course we get some upscale on the new platform. Tests ran twice to ensure congruence, showing "best" result.

Left/Top = Z270 platform / Right/Bottom = Z790 platform

Such an upscale is much less obvious however with the 970 EVO, while there was still a trend with RND4K getting weird results, however it seems to have somewhat shifted in the Mix instead of the Reads for most runs EXCEPT in RND4K Q1T1 where I COULD NOT regularize Reads/Mix at all neither reciprocate Mix especially: either Reads would fail, or sometimes both failed, and it shifted like that on all (nearly a dozen) runs I did. So I kept two results there to image it.

So the Def Peak run showed a slight improvement in RND4K Q8T1 with the 970 EVO, although it's not very substantial. For the rest of the tests, it's all pretty much flatlining.


For an extra effort, I compared both drives as well in ATTO, using the "per default" test, then changing the options until I found the ATTO test that was the most "excruciating" for both drives, which is to Bypass Write Cache, adding Verify Data, which in turns gives the option to change the test pattern to "Random" (per default would be "00000000" which seems to be the same thing than 0Fill in CDM) and to change the Queue Depth to "1" instead of "4".

Left/Top is the Optane, Right/Bottom is the 970 EVO.

Where Optane is superior until 8K blocks, both drives about equalizing at 16K, and the 970 EVO being superior from 32K onwards.

And THAT'S IT. No more testing, no more discussion of testing, I am done with adding new data to this thread.

Thanks all for patience and participation.

 

[Ruru]

I totally forgot about Optane. I almost got a 32GB(?) drive back in the day when I had a Z270 board for a while.

 

[QcRMoD]

I totally forgot about Optane. I almost got a 32GB(?) drive back in the day when I had a Z270 board for a while.

Most of the entire world seems to have forgotten about it.

 

[Ruru]

Most of the entire world seems to have forgotten about it.

You're not wrong, as I can't remember the use of those small drives... cache?

 

[QcRMoD]

You're not wrong, as I can't remember the use of those small drives... cache?

Yep they were mostly cache drives back then, and that size is surely one.

 

[dgianstefani]

Most of the entire world seems to have forgotten about it.

I hope this continues so I can get new stock for new builds.

You're not wrong, as I can't remember the use of those small drives... cache?

3.6 TB options exist.

Hope you have a spare kidney.

 

[LabRat 891]

I hope this continues so I can get new stock for new builds.


3.6 TB options exist.

Hope you have a spare kidney.

I *almost* pulled the trigger on a 1.9TB 905P for <$400. Realized that I have 'enough' to 'play with' as-is, and if I'm lucky can score something better on the used market, later.

 

[QcRMoD]

I *almost* pulled the trigger on a 1.9TB 905P for <$400. Realized that I have 'enough' to 'play with' as-is, and if I'm lucky can score something better on the used market, later.

There barely was ever an enthusiast market for it and you’re hoping for aftermarket.

I’d chase down companies foreclosure. Perhaps you might get an entire server with a 5800X in it for 905P prices…

 

[mesaprotector]

The 3.2 TB is listed on Intel's website, but I've never seen it for sale anywhere. The 1.6TB exists and will only cost half a kidney.

Prices for the p4800/5800x have actually gone up. I don't think they're ever going to be affordable. I've never liked the idea of buying a used SSD... though for Optane I guess it makes sense.
If I ever become an eccentric billionaire the first thing I'd do is bribe Intel to make PCIe 6.0 M.2/U.2 drives with that technology, and to market them heavily this time around.

The smaller size of the 800p/p1600x is 58GB, which would be enough for an OS drive and nothing else. The 118GB version is in theory big enough for a whole system. I used to dual-boot on a 256GB split down the middle, so I know! But not super practical. I have all my pictures, videos, and music on a separate larger drive. while applications and everything else, including a 16GB cache, is on the Optane. I still have 39GB free, so I don't anticipate running out of space any time soon.

If you have an unused drive slot and know how to do a reinstall, I'd just jump on these - you get a more responsive system for $30! Just know not to use them for stuff that requires fast sequential speeds. Resuming from hibernation is sloowww.

 

[dgianstefani]

Most of the 5800X on ebay are new. They're going to continue going up in price because literally still the best, by far, and no longer in production.

Supply and demand.

 

[Fokker]

Most of the entire world seems to have forgotten about it.

Intel didn't really put a whole lot of effort into advertising it from what I've seen. I've almost only seen it used in the wild as a buffer cache for HDD operating in a disk array.

 

[QcRMoD]

Most of the 5800X on ebay are new. They're going to continue going up in price because literally still the best, by far, and no longer in production.

Supply and demand.

But as far as this is true in prices it's still idealistic in demand.

Proof is most of these are still there.

Even now as supplies dwindle and prices go up... like what 3 years after they were called extinct... and warranty might really fire back onto you.

For most people it'll be a choice of... almost no perceived benefits for price and capacity and overall performance. Overall performance DOES NOT especially underlines Optane even. I've shown a 3.0x4 regular nVME to be on top of most Optane for anything over Q8T1 Rnd4K... let alone marketing pushing most of the market real far from Optane.

RND4K impresses about no one. OS for most, as I've already pointed out, is only a gateway into stuff where capacity or Sequential or more than 16KB blocks or RND4K Q1T1.... really matters... or just more aptly put... where anything CPU/RAM/GPU is much more of a driving choice.

I've posted all this cause I respect Optane real target. But it's not like I am truly just encouraging ANY instance of needs to go there. Most WON'T CARE AND WON'T SEE A DIFFERENCE.

Proof of that is SINCE THE TIME they were called extinct, if any acknolwedged gain would have been felt there, it'd be a long time all of these after-extinction stocks would be gone.

Yet they persist, and as you say, are still available in a new state.

 

[Selaya]

There Is Nothing Wrong With Optane And 3D Xpoint (NASDAQ:INTC)

Both Intel and Micron have dropped their pursuit of 3D XPoint / Optane. There is nothing wrong with the well-received technology itself. Read more here.

seekingalpha.com

^ why optane prices are skyrocketing fast

im glad i bought a 100gb optane on the cheap-ish when i first built my computer; i was tempted for about 5 minutes to wait it out for 2nd gen pcie 4.0 ones until it hit me that i'd never be able to afford one
turns out my instincts were spot-on

 

[QcRMoD]

There Is Nothing Wrong With Optane And 3D Xpoint (NASDAQ:INTC)

Both Intel and Micron have dropped their pursuit of 3D XPoint / Optane. There is nothing wrong with the well-received technology itself. Read more here.

seekingalpha.com

^ why optane prices are skyrocketing fast

im glad i bought a 100gb optane on the cheap-ish when i first built my computer; i was tempted for about 5 minutes to wait it out for 2nd gen pcie 4.0 ones until it hit me that i'd never be able to afford one
turns out my instincts were spot-on

Nothing wrong indeed. Optane had the future of SSDs right.

But now it's exctinct and beyond capabilities in RND4K Q1T1, what really is there? Outlandish $/GB is there. NO ONE CARES because it's not something readily felt or needed towards capacity and $/TB.

And that IS because OS and primary apps are a gateway into something that will just NOT USE an Optane capabilities - well, at least, not nearly as much as synthetic tests would show.

Let me just give everyone an example and propose that they shall prove me wrong:

I've just transfered my best friend from an i5 4570 DDR3 on an old HyperX Fury Sata 120GB drive to a 12600K/SN850X. Then I've restarted his old PC in Win 11 still on that old POS Sata SSD 120GB.

I've built a friend of my friend a 13600K/Z790 PC DDR5 a year ago with a SN770.

I've reinstalled my old PC on an Optane 1600X in Win 10 and Win 11 compared to doing it with an old 970 EVO.

I've then upgrade my old PC with Z790 i5 14600K just the last week on Win 11.

Guess which system is BY FAR THE FASTEST TO EITHER BOOT, REBOOT, OR WAKE FROM SLEEP?

It's the 4670 DDR3 with the old Sata SSD 120GB... and I'm not talking quantum differences... I'm talking HUGELY speedier.

DDR4 and DDR5 and newer CPUs if not RAM.... they've so much latency everywhere even with Optane 1600X... they FEEL SLOW and ARE INDEED SLOW in all of those three operations that will mostly define Everyone feeling of a speedy OS. By FAR.... it's like 250% faster on a Sata SSD and DDR3.

And then what we're talking about with Optane is faster background/launching apps benefits? And as rebooting/booting/waking up is concerned, an old DDR/SSD Sata feels speedier.

Yeah do not convince me.... I got the Optane.

Just convince me to convince ANYONE OF MY REGULAR CUSTOMERS, let alone FRIENDS, that they should purchase into it.

 

[dgianstefani]

But as far as this is true in prices it's still idealistic in demand.

Proof is most of these are still there.

Even now as supplies dwindle and prices go up... like what 3 years after they were called extinct... and warranty might really fire back onto you.

For most people it'll be a choice of... almost no perceived benefits for price and capacity and overall performance. Overall performance DOES NOT especially underlines Optane even. I've shown a 3.0x4 regular nVME to be on top of most Optane for anything over Q8T1 Rnd4K... let alone marketing pushing most of the market real far from Optane.

RND4K impresses about no one. OS for most, as I've already pointed out, is only a gateway into stuff where capacity or Sequential or more than 16KB blocks or RND4K Q1T1.... really matters... or just more aptly put... where anything CPU/RAM/GPU is much more of a driving choice.

I've posted all this cause I respect Optane real target. But it's not like I am truly just encouraging ANY instance of needs to go there. Most WON'T CARE AND WON'T SEE A DIFFERENCE.

Proof of that is SINCE THE TIME they were called extinct, if any acknolwedged gain would have been felt there, it'd be a long time all of these after-extinction stocks would be gone.

Yet they persist, and as you say, are still available in a new state.

You've got a £60 P1600X running at gen 3 which still gives gen 5 NAND flash drives a run for their money in real world performance (the average user doesn't spend their time copying sequential movies from one drive to another), yet it's really nowhere near comparable to a £3-5k P5800X which is approximately three to six times faster than the best flash you can buy, and 10-20x faster in certain workloads, with latency a full order of magnitude lower, and endurance two orders of magnitude higher, so much so that it can literally be used as RAM. There some SLC enterprise drives, and low volume specialist drives that compete in some limited ways performance wise, but still don't come close to the endurance of Optane, nor it's consistency regardless of workload.

Even a maxed out capacity 905p doesn't come close.

As a reminder, here's a nice flat line, with the conventional drives all the way at the bottom, SLC caches full and algorithms frantically trying to keep up.

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.

View attachment 337395

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.

The people who use Optane enterprise drives can take advantage of this, the average consumer can't. Not really. Not on consumer platforms with PCIe lanes that share DMI bandwidth etc. Regardless of how much benchmarking you do. There's perceived OS and application speed and peace of mind that your drive won't fail from cell wear, but you won't see your FPS number double, that's not how this works.

You can't make accurate assumptions about other models within a product line no matter how much testing you do of the entry level model.

The actual enterprise class drives are still around (at twice the price they were a few years ago) because as supply dwindles, demand goes up, and so does price, so they're still selling even at inflated prices, just look at sold listings. People who know about Optane have already bought theirs, and they don't exactly break or wear out from usage like a typical drive, so those people aren't needing replacements either.

Serve the home etc and level one techs can give you more details on what I'm talking about. They tend to snap up the fire sales of Optane when they happen (rarely). Right now most of the drives are coming from China for some reason, I suspect there's little to no new stock left in the West.

There is no warranty for EOL products, not that you need it with this class of product.

 

[QcRMoD]

And BTW it's not the FIRST time I've either restarted or switch Windows on an older DDR3 / Older SSD Sata system. They are ALWAYS much speedier than with D4/D5 systems with any nVME in there.

You've got a £60 P1600X running at gen 3 which still gives gen 5 NAND flash drives a run for their money in real world performance (the average user doesn't spend their time copying sequential movies from one drive to another), yet it's really nowhere near comparable to a £3-5k P5800X which is approximately three to six times faster than the best flash you can buy, and 10-20x faster in certain workloads, with latency a full order of magnitude lower, and endurance two orders of magnitude higher, so much so that it can literally be used as RAM. There some SLC enterprise drives, and low volume specialist drives that compete in some limited ways performance wise, but still don't come close to the endurance of Optane, nor it's consistency regardless of workload.

Even a maxed out capacity 905p doesn't come close.

As a reminder, here's a nice flat line, with the conventional drives all the way at the bottom, SLC caches full and algorithms frantically trying to keep up.


The people who use Optane enterprise drives can take advantage of this, the average consumer can't. Not really. Not on consumer platforms with PCIe lanes that share DMI bandwidth etc. Regardless of how much benchmarking you do. There's perceived OS and application speed and peace of mind that your drive won't fail from cell wear, but you won't see your FPS number double, that's not how this works.

You can't make accurate assumptions about other models within a product line no matter how much testing you do of the entry level model.

The actual enterprise class drives are still around (at twice the price they were a few years ago) because as supply dwindles, demand goes up, and so does price, so they're still selling even at inflated prices, just look at sold listings. People who know about Optane have already bought theirs, and they don't exactly break or wear out from usage like a typical drive, so those people aren't needing replacements either.

Serve the home etc and level one techs can give you more details on what I'm talking about. They tend to snap up the fire sales of Optane when they happen (rarely). Right now most of the drives are coming from China for some reason, I suspect there's little to no new stock left in the West.

There is no warranty for EOL products, not that you need it with this class of product.

I've never said the theory and some of the practice is untrue my friend. I've bought the Optane because of you.

I'm just giving a reality perspective. Either 970 EVO, Optane 1600X on an old system or a brand new one, either SN770 or SN850X on a up to date one, can't even manage to boot/reboot or wake up faster than when we had SSD satas and DDR3 limited systems.

An entirely different way of looking at HOW IT WILL FEEL for MOST USERS.

Then add my argument since THE BEGINNING OF TESTING THE OPTANE: For most people, the OS and all of what Optane does best is only a brief gateway into something else that will require more than an Optane to show the best results: better CPU, better RAM, better GPU.

In fact you should thank me.... if my truly realistic opinions is heard, you have a chance to score any Optane drive without further hype AT ALL added to it. And I'd do it too, cause I'm as close an to the OS in use enough that I've SEEN benefits already.

Aaaaan just as much as I say this, I'm even LESS giving any nVME drive any benefit at all unless they're running on an outdated DDR3 system.

 

[dgianstefani]

And BTW it's not the FIRST time I've either restarted or switch Windows on an older DDR3 / Older SSD Sata system. They are ALWAYS much speedier than with D4/D5 systems with any nVME in there.


I've never said the theory and some of the practice is untrue my friend. I've bought the Optane because of you.

I'm just giving a reality perspective.

I don't think you can come to conclusions on reality of a different product by reviewing one product.

Your experience of older systems being speedier is simply because their architectures are slower, but more basic, less feature ridden but with lower latency and simpler software, as you've acknowledged. Complex software and wide architectures don't do well at feeling snappy.

Zen 4 needs a minute to train the RAM each boot, this is completely unrelated to whatever the SSD drive is doing, in fact the drive isn't even loaded at that point.

You're right that it's hard to stress the actual performance of Optane in typical consumer usage, but that's kind of the point of the drive, you won't be bottlenecked by it, nor will it wear out within the useful life of the computer it's connected to.

Where this stuff shines is in server applications, hence that's the target market, and you don't see these drives sold to consumers, not really.

Intel didn't really put a whole lot of effort into advertising it from what I've seen. I've almost only seen it used in the wild as a buffer cache for HDD operating in a disk array.

It was never meant for consumers, and enterprise were well aware of it and used it where it offered benefits (high stress servers and RAM replacements). There was some limited marketing re star citizen, and half hearted bundling with prebuilts as "caching" to try and normalise the tech, but it never caught on because they were trying to make an expensive product appeal to budget users, which doesn't make sense. "Cheap HDD/SSD can be fast with Optane" isn't exciting or revolutionary, and it undersold the tech. Should have gone premium to the hilt and kept making replacements for the 905p. Sadly the P5800X series are the only second generation Optane on faster PCIe 4 interface etc.

The 3.2 TB is listed on Intel's website, but I've never seen it for sale anywhere. The 1.6TB exists and will only cost half a kidney.

Prices for the p4800/5800x have actually gone up. I don't think they're ever going to be affordable. I've never liked the idea of buying a used SSD... though for Optane I guess it makes sense.
If I ever become an eccentric billionaire the first thing I'd do is bribe Intel to make PCIe 6.0 M.2/U.2 drives with that technology, and to market them heavily this time around.

The smaller size of the 800p/p1600x is 58GB, which would be enough for an OS drive and nothing else. The 118GB version is in theory big enough for a whole system. I used to dual-boot on a 256GB split down the middle, so I know! But not super practical. I have all my pictures, videos, and music on a separate larger drive. while applications and everything else, including a 16GB cache, is on the Optane. I still have 39GB free, so I don't anticipate running out of space any time soon.

If you have an unused drive slot and know how to do a reinstall, I'd just jump on these - you get a more responsive system for $30! Just know not to use them for stuff that requires fast sequential speeds. Resuming from hibernation is sloowww.

That's the perk. £60 for 120 GB of responsiveness you can't get from any other component upgrade (besides perhaps a RAM drive, which is massively impractical), consistent performance and endurance where you can simply forget about wear. You're not going to find benchmarks that show entry level Optane beating everything else by 10x, nor will a faster drive make RAM train instantly or your PC boot twice as fast (because drive speed isn't the issue there), but it is what it is.

There's also the factor that modern applications and games are literally designed around slow storage speeds, they're made to cache things in RAM. DirectStorage and other initiatives attempt to fix this, but they haven't really taken off, since flash storage hasn't really gotten much faster in a meaningful way in the last ten years, so there's little point investing in rewriting code. Better to just take advantage of RAM. The PS5 literally processes it's drive faster than any consumer PC, because it has a dedicated accelerator to do so, whereas every PC processes data through the general purpose CPU cores. That's how little focus there's been on improving storage speed. When something better than flash happens, whether that's Optane or something else, I suspect this will change. Maybe the rise of LLM, but that's really just led to usage of HBM volatile memory (again, a variant of RAM), not faster flash or other non volatile memory (parallelism doesn't count).