Intel Optane 900P SSD

so.. 800p is confirmed with just x2 lanes, not x4. and looking at the pcb/ number of chips, i'd dare say its only 2-3 channels hence the x2 lanes, its probably going to get capped at around 1500-1600 MB/s max and that is being very optimistic.

sequential side of thing is acceptable, we know 4k random read is going to be fast, all thats left is 4k random write.

 

They have probably designed this one for use in thin and flimsy with wimpy cooling. I wonder what the power consumption numbers will be. Haven’t seen much from the small cashe disks. Small steps... Baby steps

 

yea, they say its now idle at 1w, which imo still need improvement. i read anand's article they mentioned a x4 optane m.2 is coming later in the year on roadmap but they didnt post a roadmap picture/info. x4 would work well, 1 SSD no raid, very simple and be done with it.

 

I wrongly listed my benchmark as a 960 pro, when it is actually an Evo, that is in my signature. Sorry for the confusion.

 

https://www.tweaktown.com/news/60582/phisons-e12-controller-featured-ces-2018/index.html
wowzer E12 controller. MLC too, giving 960 pro a run for it's money

 

Nice write speeds. The Sammys are out now, so there's one problem. Still intriguing. We'll have to wait to see on the release and actual retail pricing.

 

i just tried a bunch of stuff torrenting while using OS on samsung SSD has some lag and lockups. torrent really destroy it because random write workload which is especially hard on TLC.

however on 900p its blazing fast. @Papusan @Phoenix

 

Nice. what is capacity? 160GB?

 

Nice. That's why I torrent on my SanDisk Extreme Pro 512GB. Be smart....be like Phoenix...

 

yea thats what i find on sandisk extreme pro is that their small files read are much better than most ssd, even the samsung pro versions. its really a shame i had at least 3 extreme pro go bad on me in a raid array that i can't trust my files to them anymore.

i must say the 4k write is a bit disappointing considering its optane memory, but there will always be new optane product so that can wait.

280gb

 

I'm using 850 evo and really torrents tax it heavily.
@ole!!! Did you OP Optane SSD? Did you deploy Optane SSDs anywhere and does it really benefit aside from NVMe drives?

 

yes and no, really towards the usage i guess. on heavy loads the 900p really starts to shine. on light loads you simply can't see much of a difference.

i have so much ramdisks, software, task schduler, settings etc all set on auto start and a bunch of garbage files with load of icons on desktop. so when i boot into desktop you can see the icons all not loading right away with 850 pro, with PM961. but with optane its almost instant, but again thats heavier loads.

game loading a tad faster, depends on the software. file copying, it destroy pretty much ALL flash based SSD except maybe the better ones like sandisk extreme pro or intel enterprise type. when i make a backup image or restore either from/to 900p, it is much quicker than NVMe because of the random read/write.

much to my surprise, file search seems only a bit faster and its a lot of random write involved no idea why.

btw when you torrent on 850 evo and say you try to start a game does it lag/hang for 2-3 sec? assuming its torrent to 850, and game loads from 850

edit: optane doesn't need OP forgot to mention that.

 

Overprovisioning beyond the drive's native amount will have minimal impact on performance for Optane. For normal consumers without a very very special need you will never ever need to do it. Damn, they are very expencive, they are small (even smaller for laptops)... Use the available place if needed. Optane drives ain't the same tech as ssd's. Not the same as a small trashy todays 120GB NVMe ssd for your OS and software. You won't manage to kill it with your use

 

On Optane, the write numbers are pure 3D XPoint performance, while write performance on NAND SSDs are buffered behind DRAM.

 

This is my CDM benchmark on Wine using 850 evo m.2 250GB. The R/W speeds are outrageous and I have no RAMDisk SW or Samsung TurboWrite tech installed on linux.

 

not sure, but that speed is only possible if ram is involved. 4k random read/write qd1 represent what ddr4 ram is capable of.

 

yes for 3d Xpoint, for nand, the dram use to be the buffer, however nowaday it does a lot more than just buffer and the actual buffer impact a lot less. nand is actually very fast, take a look at sandforce SSDs they are dram less and still plenty fast, loved the SLC + sandforce SSD blazing fast in the past.

for TLC, im sure dram PLUS their SLC section are doing bunch of work thats why people that buys evo are blinded by the numbers. on MLC its a lot better and more so represent more closely to actual flash performance, or controller performance whichever is the lowest.

something i'd like to add is that, 900p has rather poor performance for files very small block size vs say sandisk extreme pro, or even samsung 960 pro. 512b, 1k, 2k 4k etc are pretty slow when against these flash SSD.

 

Not sure about. Maybe its cached readings since I ran it on WINE.

 

Instead of Crystal Disk Mark under Wine, how about a real-world test to get your numbers?

Try grabbing some large .tar, .iso files or dd from a DVD. Next copy location to location, and time it. What 'ballpark' i/o throughput are you seeing? IIRC, I'm seeing something like ~2.5 to 2.65 GB/s on the Samsung 960s using large file copies. However, this was copying with two disks... copying disk to disk (going from nvme01 to nvme02).

 

I'll check that. IIRC, Copying W10 ISO from PM951 to 850 EVO gave me write speeds of 490MB/s. Will check it thoroughly.

 

if you're doing this with window copy system, try doing it with fast copy. windows like to put stuff into ram since forever then back onto SSD, where as fast copy doesnt really use ram at all so you can see way more details. you can choose buffer size, alt stream etc etc, its really good.

you can see the Q1 sequential read/write performance for device you are copying it to, how many small files sequential copies affect which drive etc, which also shows sandisk extreme pro SATA SSD is much superior to even NVMe 960 pro when comes to lots of small block size files.

 

That was copying the .iso within Linux? What are the specs of your 850 EVO? Do those numbers match up? I would trust those numbers rather than what Crystal Disk Mark is reporting while running under WINE on Linux.

My numbers are from when I was moving some VMs within Linux Mint. I was seeing close to the spec write performance of the Samsung 960 NVMe drives.

 

Great post @Dr. AMK.

@tilleroftheearth, I know you'd rather have text instead of a video, but some of the reporting in this clip seems like it applies to your unorthodox work environment. I think the endurance part of the Optane 900P may be of interest to you.

 

im not too sure i don't use Linux, but i do know it varies between OS. example, consumer windows 7/8 would use ~2GB of ram. so if you were to copy from HDD/HDD a 10GB zip file you'll see your ram usage goes up max up to 2GB, and when it says copy is done you take a look at your laptop it still loads and HDD lights are flashing copying from ram back down to HDD. if you were to shutdown during this phase your file likely will corrupt.

im on window server OS, that 2GB limit becomes like 512GB so it became unusable for me to copy big files and i use fast copy, doesn't go through ram at all, much better and faster than even terracopy. it also shows a lot of info too.

850 evo will not have those numbers for sure unless ram is involved, since SATA SSD max around 550MB/s, and 4k random should be around 40-50MB/s max.

 

@ole!!! - I think you've misinterpreted.

My post was in response to this thread - http://forum.notebookreview.com/threads/intel-optane-900p-ssd.810226/page-12#post-10670068 by @Vasudev who was running Crystal Disk Mark under Wine (which purports Wine Is Not a windows Emulator) and had some fantastic bench numbers on the 850 EVO.

 

I'm pretty happy with the NVMe numbers I'm seeing (when I boot into native Windows) See the spoiler of the benches in this thread - http://forum.notebookreview.com/threads/mighty-muscular-mini-itx-build.812322/#post-10656108

 

disable DMI ASPM and SSD PCIE ASPM, turn off cstate and do another test, you'll get a massive jump in score especially in the 4k.

 

ASPM settings... Is this something to do in Windows or something to look for in my BIOS?


---

Edit, FWIW:

The power plan in Windows 10 has PCI Express -> Link State Power Management -> Setting: Off

In UEFI/BIOS, the following are already configured:
PCIE ASPM Support <Disabled>
PCH PCIE ASPM Support <Disabled>
PCH DMI ASPM Support <Disabled>

 

its bios, but looks like you've found them already. if they already are disabled then only boost would be cstate off but that would mean high clock all the time lol not worth the power. rest are difference between OS, linux vs windows i guess.

 

writing small files as much as possible at lowest QD possible (QD 2 with ATTO). notice how first 2 picture both are 961 or 960 EVO oem.. but 2nd picture shows pure trash performance vs the first. in comparison look at sandisk extreme pro.

its likely read is not stored in the SLC layer so that pure junk read performance is likely true performance of TLC flash at small block size file level. which also explains why my hyperx 3.0 is pure sh!t too, just like many people reviewed the drive also mentioned slow files copy into it is slow as hell.

 

What is the PM961 page size? If it is 4K or bett than that may explain the read performance, but other than Atto's .5K to 4K, the rest of the numbers on the Samsung look good compared to the other two.

 

this is all single drive, if you're talking about default format size its all 4k, same with Hyperx 3.0 and sandisk extreme pro. if you mean the page size by samsung its likely 16k or 32k, this is something im not too familiar with however my understanding is that increasing this page size via firmware on the flash would increase performance, but lower endurance.

 

No. I'm talking about the size of the resting place where the SSD controller stores actual data within its group of NAND cells.

Do you know if the disks were secure erased before the ATTO test? Furthermore, do you enough about the ATTO test's timings? It has to put files on the SSD, correct? If pages are dirty, there's a write amplification problem to worry about. Does ATTO take things like that into account?

[ Note, this is a bit dated (as it won't cover some of the newer technology like TLC, 3D vertical, etc.) , but for anyone wanting a primer behind on the technology behind SSDs I *strongly* recommend this entire article, but one can start here - https://www.anandtech.com/show/2738/5 thru https://www.anandtech.com/show/2738/14 ]


What is strange, if you look at WRITE speeds the PM961 has better numbers hands down. It is READs that are interesting, but only until the file size reaches 8K. At that point, PM961 equals and then gets better performance numbers. My post above questions if it is something to do with the NAND page size, the ATTO test, or perhaps testing against a non-secure erased SSD.
-------

Read Speeds

PM961
.5K = 527 MB/s
1K = 1062 MB/s
2K = 2103 MB/s
4K = 4161 MB/s
8K = 38,060 MB/s
16K = 271,302 MB/s
32K = 323,649 MB/s

Sandisk
.5K = 14,123 MB/s
1K = 19,726 MB/s
2K = 22,740 MB/s
4K = 26,820 MB/s
8K = 43,892 MB/s
16K = 91,799 MB/s
32K = 145,244 MB/s

Kingston
.5K = 3170 MB/s
1K = 5881 MB/s
2K = 9633 MB/s
4K = 13,811 MB/s
8K = 18,071 MB/s
16K = 20,103 MB/s
32K = 30,736 MB/s

 

Actually I came here to post that video, to show that so far Optane in this form is worthless...and SATA is still "fast enough" for real-time 99% use.

 

I was hoping @tilleroftheearth would've had some kind of comment.

 

He seems to have taken an NBR "New Years Resolution"...he hasn't posted here in 2018... yet.

 

the issue is this performance can be duplicated on clean drive or dirty drive. needless to say we shouldnt be testing it on clean drive let alone secure erase, because drive are meant to be used so it should always be tested with files in it.

PCIE has lower latency, not to mention PM961 is TLC drive so it has SLC caching. so once SLC filled up TLC write is only 600MB/s and thats with multiple channel TLC writing at the same time (kinda similar raid 0 with SSD controller being the raid controller, more channels flash = faster sequential output, also another reason why optane memory m.2 is only 300MB/s because only 1-2 channels where as optane can max at 7 with 900p)

it should go like this if my guess is correct (tho i have no idea if this is how ATTO works)
ATTO --> SLC --> many TLCs flash for write
many TLCs --> SLC/no SLC --> ATTO for read

wrhite test: 32M OR 64K writes into SLC so write performance same, however it goes into SLC to TLC soon enough that its no longer stored within SLC cache, so that means when read test happens, it wont come from SLC caching, it'll come straight from TLC.

read test: assuming the page size is 16K, then a small 64K would be broken into 4 * 16K which is either 4 channels at QD1 or 2 channels for QD2 (and so on) dependent on the firmware how controller reads it's flash. assuming it is 2 for QD2 for 64K, then 32M would 512 * 64K which would put this to multiple channel and multiple QD, thus increasing the speed drastically (saying QD here isnt the right term but idea is the similar)

perhaps this is the reason behind it and shows true TLC performance, or at least samsung's TLC performance.

thats like saying intel 2 core at 5ghz is faster than ryzen 8 cores 4ghz because a software only uses 2 cores. Optane hardcore destroys any other flash based SSD, at this point its software not taking full advantage of it, heck software isnt even taking full advantage of traditional HDD/SSD.

if i were to recommend to a friend yea i won't bother telling them to buy optane, a PCIE based SSD is good enough for them.

 

copying almost 70k files, only 120MB in size. 1/3rd of files less than 1k, 1/3rd is 3KB and 1/3rd is 5KB. writing to Z:\ which is a ramdisk to remove any write bottleneck so read speed will be the bottleneck here per device. Fastcopy 2.11 use 1 thread read, 1 thread write mode, bypass OS cache so its not device writes into OS cache (ram) then into ram disk again.

 

Can you zip up that file bundle and make it available somewhere?

 

i cleared them, but you can just find small files within windows either compressible or in compressible shouldn't matter to the SSDs nowadays. copy it to folder and redo that like 100 times, then select the 100 copy it another 100 times u'll get 10k files.

edit: i just made another one uploaded to mega. donno how to do one without encryption key its pretty annoying. https://mega.nz/#!yAo1AbgC!dnmAjgJDmJysMR1ixPOvT_JcI44Er6SaNpSyQ60mpxM

 

That's alright. I was hoping to run the same test using the same exact files on my end to get an idea of comparisons. Maybe another time.

 

Great! Can you get post of some screen shots of FastCopy? Should I use the defaults on my end? I'm seeing a 256MB buffer.

 

specific settings to watch out for cause they dont show in my screenshot, i'll test it with fast copy later tonight. in my screenshot is old version 2.11, if you're using 3.4+ then use these settings i'll be testing with same settings.

defaults
- buffer 256 MB
- Alt Stream uncheck
- ACL uncheck
- verify uncheck
- ETA finishTime uncheck
speed control at full speed

I/o settings
- Max I/O size 1
- num of overlap I/O issue 1
- make sure MinSectorSize box unchecked
- OS cache settings, change both on NTFS and FAT box to 1. (default is 64 on NTFS and 128 on FAT).
- uncheck box directory fetch

rest don't really matter.

 

OK... Here's a copy from the Samsung 960 Pro NVMe (512GB) to a Ram Disk.

Spoiler: Complete Fast32Copy test

What Ram Disk should I use? I'm using IMDisk, and with those settings, I see the 'WRITES' to the ramdisk are the bottleneck. Reads from the Samsung 960 Pro are done 14s into the test, but there are still 72MB of writes to the ramdisk needed to finish.

Spoiler: Reads done (14s into test) during a fresh Fast32Copy test

 

check your settings defaults and I/O settings with mine.
I use primo ramdisk, cluster size is 4k, v5.6.0

 

@jclausius how did the test go with those settings? btw i use logical, scsi disk. not direct I/O.

 

Sorry didn't get a chance to mess with it. I was working on the Linux side of things this weekend. Something is definitely wrong as I don't think the NVMe should perform that poorly on Reads and the Ram Disk should outpace that. Not sure what gives, but this is not a priority to fix as I'm not really using Win 10 much these days.

That reminds me. What OS was this tested under? I'm running Win 10 pro with no tuning after plain driver installs. What clocks? I run just stock.

 

• Intel Optane SSD 900P 280GB Hardwareluxx (de)

 

Will Phoenix buy this drive for his new desktop?