News Micron Finally Rolls 3D Xpoint SSD: X100 Billed as 'World's Fastest' with 2.5 Million IOPS and 9 GBps

Impressive darn specs!

Get this to consumer drives (and prices) ASAP! (Folks bottlenecked by storage speeds might pay $400 or even $500-$600 for a 1 TB drive that fast, but, most consumers are not going to pay the ridiculous $2k-$3k that some of the 'DataCenter'-intended cards go for...

Trying already to imagine installing Windows in....2 minutes! :)
 
it has an'8 pin power' input? as in a PCI-e 8 pin like that typically connected to a GPU?

Hard to imagine the 75 watts available from a PCI-e slot not being sufficient, but,..I guess we hope to know in 4 years when this hits desktop at a decent price point. :)
 
Reactions: bit_user

DavidC1

Distinguished
May 18, 2006
391
6
18,785
0
Get this to consumer drives (and prices) ASAP! (Folks bottlenecked by storage speeds might pay $400 or even $500-$600 for a 1 TB drive that fast, but, most consumers are not going to pay the ridiculous $2k-$3k that some of the 'DataCenter'-intended cards go for...
It won't.

The volume is too low to be that low priced. Some reported Intel barely making profits on the enthusiast Optane 905P SSDs.

it has an'8 pin power' input? as in a PCI-e 8 pin like that typically connected to a GPU?
This makes sense. Current 3D Xpoint requires lots of power to reach high bandwidth. It's probably the reason Intel's Optane SSDs are relatively low for sequential specs, so it doesn't use so much power.
 

seanwebster

Contributing Writer
Editor
Aug 30, 2018
89
20
545
3
it has an'8 pin power' input? as in a PCI-e 8 pin like that typically connected to a GPU?

Hard to imagine the 75 watts available from a PCI-e slot not being sufficient, but,..I guess we hope to know in 4 years when this hits desktop at a decent price point. :)
SSDs can only pull up to 25W from a PCIe slot. It needs that 8-pin to supplement anything draw more than that.
 

Brian_R170

Honorable
Jun 24, 2014
288
2
10,785
0
Get this to consumer drives (and prices) ASAP! (Folks bottlenecked by storage speeds might pay $400 or even $500-$600 for a 1 TB drive that fast, but, most consumers are not going to pay the ridiculous $2k-$3k that some of the 'DataCenter'-intended cards go for...
Don't hold your breath. Intel Optane SSDs are over $4/GB for datacenter (P4800X) and over $1/GB for consumer (905P). I would expect the higher-performing Micron drive to be even more expensive. It could be well over 5 years before you see a consumer version anything close to $0.50/GB.
 

derekullo

Distinguished
Jan 25, 2009
2,266
17
20,465
241
Very impressive.

The only thing I am wondering is why they stopped at 9 gigabytes per second?

The maximum bandwidth of PCI-E 3.0 x16 is 15.75 Gigabytes per second.

The 1.58% overhead from 128b/130b is negligible, 15.49 Gigabytes per second.

The PCIE 3.0 x4 Samsung 970 Pro has a 3500MB/s read which is 90% of peak bandwidth of PCI-E 3.0 x4. (3500/3877 Adjusted for overhead)

The Micro X100 has a 9000MB/s read and write which is 58% of peak bandwidth of PCIE 3.0 x16. (9000/15490 Adjusted for overhead)

I'm assuming if they could have they would have since their target audience has deep pockets and for the bragging rights.

The only reason I can think of would be heat since it does have an 8 pin in addition to the PCI-E slot.


PCI-E 4.0 x4 is looking good if this ever comes in that flavor!!! 7.88 GB/s max bandwidth for that.
 

bit_user

Splendid
Herald
Similar prices to what SATA SSD's were originally.
Then NVMe.

This too will come down eventually.
Well, it was a given that NVMe would come down, because the only real difference between that and SATA is the controller. Plus, M.2 drives avoid the added cost of a case. Yet, the fastest NAND-based NVMe drives will continue to command a premium, since they'll use lower-density NAND to get better performance.

However, you're assuming that 3D XPoint will follow the same price curve as NAND, and I don't know if that's true. If we look at how NAND got cheap, it's because:
  1. Manufacturing node shrinks
  2. Packing multiple bits (now up to 16, with some talking about 32) per cell
  3. 3D stacking
As for #1, presumably 3D XPoint is made on a somewhat comparable node to recent NAND. Maybe previous-gen, but suffice to say that it's not starting from way behind, so it doesn't have much room to catch up, here.

Regarding #2, I don't know if 3D XPoint can manage the same trick, but it doesn't sound like it.

So, that leaves #3. We'll have to see how well it can live up to its name and the supposed promise of being 3D, but I think the first gen chips were planar.

In conclusion, if they can't pack multiple bits per cell, then it's going to need density improvements from stacking and areal density, in order to compete with NAND. I don't know how much potential there is for either, but it's probably not a stretch to say it's not likely ever to compete with NAND on price. What we'll probably see is a world where HDDs continue offer the best GB/$, 3D XPoint offers the best performance, and NAND sits somewhere in between.
 
Aug 8, 2019
36
5
35
0
Looks good on paper but hardly matters for most users.....The problem is never with sequential read/write. Its nature of I/O.... Because most of our daily I/O consist of numerous small files at very low queue depths (hardly more than 5).

Another is windows itself. How I/O is performed. Windows reads file by file (serial form).... So transfer rate becomes very low when you have numerous small files. This needs to change before we can see big improvement.

We have already seen performance with 905P....... In real life usage, performance boost over normal NVMe SSD is very small.
 

bit_user

Splendid
Herald
Looks good on paper but hardly matters for most users.....The problem is never with sequential read/write. Its nature of I/O.... Because most of our daily I/O consist of numerous small files at very low queue depths (hardly more than 5).
Low-QD IOPS is actually where 3D XPoint is unparalleled.

Another is windows itself. How I/O is performed. Windows reads file by file (serial form).... So transfer rate becomes very low when you have numerous small files. This needs to change before we can see big improvement.
Apps that will benefit use I/O Completion Ports or other means of asynchronous I/O (Linux has its own flavor - Posix AIO is poorly designed and the Linux implementation is even worse).

However, you don't only have one app accessing storage at a time, and that's where you can easily see the benefits of higher QD. Still not common, for desktop users, but then this product isn't aimed at that market.

We have already seen performance with 905P....... In real life usage, performance boost over normal NVMe SSD is very small.
Using conventional benchmarks, but has anyone ever managed to quantify "responsiveness" in a real-world benchmark?
 

DavidC1

Distinguished
May 18, 2006
391
6
18,785
0
Using conventional benchmarks, but has anyone ever managed to quantify "responsiveness" in a real-world benchmark?
It's even worse in real-world load testing. Optane loads minimally faster.

The only thing I am wondering is why they stopped at 9 gigabytes per second?
It has to be heat. If you look at the M.2 version of P4801X datacenter Optane drives, the lower capacity version lowers power but has significantly lower speeds too.

Their DC PMM modules are capped at 18W, and so are their SSDs, and they both achieve similar ~2GB/s.

Research papers are out there saying PCM memory possibly using more power with writes. It looks like its true in reality too.
 

seanwebster

Contributing Writer
Editor
Aug 30, 2018
89
20
545
3
That applies only to M.2 slots. The article is obviously concerning a full-sized x16 PCIe card, so it can get up to 75 W from that slot.
Yes, it is full length, but a company rep at the event stated it could only pull 25W from the slot, that is why the additional connector was added in. "An Enterprise PCIe SSD(s) shall get all its power from the 12V pins" and the max power limit is 25W. It is part of Enterprise SSD spec. See page 37 here: http://www.ssdformfactor.org/docs/SSD_Form_Factor_Version1_a.pdf

Although, I've seen some drives that are upwards of 35-55W without the need for an additional power input on an x8 connection. I've reached out for clarification, but they are very quite on specifics at this time.

M.2 is lower wattage ~7 - 8.25W - See page 3 here: https://www.flashmemorysummit.com/Proceedings2019/08-07-Wednesday/20190807_NVME-201-1_Chen.pdf
 
Last edited:
Reactions: bit_user

MRFS

Distinguished
Dec 13, 2008
1,333
0
19,360
40
Patrick Kennedy asked a good question in his newsletter: how does this AIC present itself to an operating system? e.g. can it be "bifurcated" like other "4x4" AICs? and, if so, would there be any performance advantage to doing so? At PCIe 4.0, the raw max bandwidth is so very high, the bottleneck may then shift back to driver and software overhead (presently 1 - (9/16) = ~43% at PCIe 3.0). Thus, thinking out loud here, if a future "X100" works at PCIe 4.0 speeds, a "4x4" option may result in processing that overhead in parallel across 4 x CPU cores. Theoretically, then, a "4x4" option may compute driver overhead faster than a default option that computes driver overhead in a single CPU core. Your thoughts are greatly appreciated here. Thanks everyone for all your excellent comments!
 

bit_user

Splendid
Herald
Yes, it is full length, but a company rep at the event stated it could only pull 25W from the slot
Thank you for the clarification. I did try to see what the limit was on the M.2 slot, but the info wasn't very easy to find. In the end, I saw another forum poster (on another site) claiming it was 25 W and concluded that's probably what you were talking about. However, perhaps they were actually confusing the spec you cited with M.2.

Anyway, thanks for following up.
 
Reactions: seanwebster

bit_user

Splendid
Herald
Patrick Kennedy asked a good question in his newsletter: how does this AIC present itself to an operating system? e.g. can it be "bifurcated" like other "4x4" AICs? and, if so, would there be any performance advantage to doing so?
A big selling point of NVMe is that you can create a large number of deep queues, specifically to get lots of transactions in flight and decouple I/O from different applications.

https://en.wikipedia.org/wiki/NVM_Express#Comparison_with_AHCI

Theoretically, then, a "4x4" option may compute driver overhead faster than a default option that computes driver overhead in a single CPU core.
If the driver is so bone-headed as to be single-threaded, then they should just lean on MS to fix that. I doubt it, though.
 
Reactions: TJ Hooker

DavidC1

Distinguished
May 18, 2006
391
6
18,785
0
Worse? What are you talking about? I don't know about this product, but the 905P is the most responsive NVMe drive you can buy!
Please think about the whole sentence, not focus on a single word.

I said its worse because benchmarks showing responsiveness tests such as with Sysmark shows 2x gains using Optane. And you already know that the 4K low QD speeds are phenomenal and traditionally believed to be what's responsible for performance in client PCs.

Yet its worse, because the gains in reality don't hold up. It's 0.1 seconds faster in TH tests. That doesn't mean Optane is bad, but due to inherent limits the gains are not worth the low capacity and absurdly high prices for most people.

Get it now?
 

ASK THE COMMUNITY

TRENDING THREADS