Intel 3D XPoint Optane SSD DC P4800X Performance Preview

[cats_Paw]

So thats on HDD side of things, how about CPU? Oh, right, I forgot, milking the consumer for 5% increments.
A good question is how much of other impressive technology they already have but know its financially better not to release yet.

 

[mdd1963]

When they can start cranking out 4 GB/sec throughput sequential, people will start buying them; at 2 GB sec sequential reads. this seems pointless compared to 960 EVO, etc...

 

[kawmic]

This is totally rediculous, the said they World be cheaper in price.

 

[Tech_TTT]

I dont get this .. these cards claim to be near RAM performance and yet still on x4 lanes ? they should be at least on x8 if not x16 lanes. and they need to make at least 8GB/s versions...

 

[mlee 2500]

Might have to rethink my initial cynicism about this product. Maybe even buy some Micron shares.

 

[Discorama]

"Neither company is particularly forthcoming with information about 3D XPoint, and we still don’t know exactly what it is, even though products are on the cusp of general availability. We’ve done plenty of investigation and surmise, like many industry analysts, that it’s some flavor of PCM (Phase Change Memory)."

OMG! Stop saying that no one knows what it is. It is not Phase Change based. It is called "Crossbar Memory"! Plenty of articles detailing the specifics on it, just read up. I can't believe people can't put 2+2 together on this... The "X" in "X-point" is Intel's spin on the "Cross" in "Crossbar Memory". Get it?!
Crossbar was supposed to be released by HP years ago along with another company called Crossbar. Intel was able to develop their own flavor of it and beat them to the punch. Samsung is also working on their own flavor as well.

 

[PaulAlcorn]

"Neither company is particularly forthcoming with information about 3D XPoint, and we still don’t know exactly what it is, even though products are on the cusp of general availability. We’ve done plenty of investigation and surmise, like many industry analysts, that it’s some flavor of PCM (Phase Change Memory)."

OMG! Stop saying that no one knows what it is. It is not Phase Change based. It is called "Crossbar Memory"! Plenty of articles detailing the specifics on it, just read up. I can't believe people can't put 2+2 together on this... The "X" in "X-point" is Intel's spin on the "Cross" in "Crossbar Memory". Get it?!
Crossbar was supposed to be released by HP years ago along with another company called Crossbar. Intel was able to develop their own flavor of it and beat them to the punch. Samsung is also working on their own flavor as well.

That's an interesting theory, but the XPoint is merely a branding of the standard crosspoint memory architecture.

IMFT arranged the material in a crosspoint architecture, which is a common alignment for memory, denoting a chessboard-style series of bitlines and wordlines. There are many different uses for crosspoint architecture, so IMFT merely added some marketing magic to an established architecture naming convention to come up with its 3D XPoint moniker.

A big reason why pretty much all of the analysts agree it must be a flavor of PCM is because of this:

Guy Blalock, the co-CEO of IMFT (from the Micron side), narrowed it down when he commented earlier this year at an event hosted by the SEMI trade group that 3D XPoint uses a chalcogenide material and an Ovonyx switch. Intel had a joint venture with Ovonyx and has licensed its technology in the past, but Micron purchased the firm for $1.3 billion in 2012 from Energy Conversion Devices (ECD). Ovonyx specializes in proprietary PCM technology (Phase-Change Memory).

right from the horse's mouth.

http://www.tomshardware.com/reviews/3d-xpoint-guide,4747-3.html

 

[nitrium]

When they can start cranking out 4 GB/sec throughput sequential, people will start buying them; at 2 GB sec sequential reads. this seems pointless compared to 960 EVO, etc...

Don't agree at all. Consumer workloads are generally not sequential. The most important metric for general users is 4K random reads at QD 1 (and most SSDs struggle to break 20 MB/sec at that task). These drives are 5-8 times faster than ANY SSD on the market for that workload. Windows, games and apps will load MUCH faster on one of these than a 960 EVO.

 

[hannibal]

These xpoints Are real speed monsters! Expensive sure but if I could afford them definitely a worth of upgrading. And just like first ssd:s these will become cheaper and faster. Do tuo remember when 16gb ssd was super expensive and was thinking if it would be wice to buy it instead of 500gb HD... Sooner or later it will be.

 

[eriko]

@ michaelzehr

I'd remove that comment if I were you - and get busy making $$$$!

 

[daglesj]

Could do with say just 4GB of this actually built onto the motherboard for cache and such.

 

[grozzie]

Whats the point of Optane anyway? It doesn't add anything that's not already available. I use a X4 Gen3 M.2 SSD for primary boot and programs, and a SATA SSD (Samsung 950 evo) for data. I still use HDD's for backup and as far as I can see, only HDD's benefit from Optane. Optain is very expensive for a very limited, not really usable capacity. The only benefit, as mentioned by others, is a cache for HDD's but even that doesn't justify the cost against performance. If you use RAID 10 with HDD's Optain might even slow data transfer rates down!

Did you even look at the benchmark results?

 

[grozzie]

Yes! I looked at supposed benchmarks but as I said - it's too expensive for little or no gain. Where would you use it? no good for a multi- VM server farmS. No good for multi-threaded processing such as 3D rendering, you can't boot off it so WHAT IS THE POINT. Oh I'm not sure if it will work with the new GEN AMD processors and motherboards.

 

[cordes85]

I have seen how cheap the Fulton x4 3.0 series AIO cards are and i am happy with those to be fair, this kind of technology looks faster than the cpu can handle lol. Ill stick with the 1.2 TB model as its fas enough as a project drive 1.2TB Intel 750 Series AIC SSD, HHHL PCIe 3.0 (x4), NVMe, 20nm MLC NAND, 2400MB/s Read, 1200MB/s Write, 440k/290k IOPs @ £650ish this new tech is ridiculously expensive, ill wait till it comes down to earth.

 

[kuocc]

Ideas from nowhere reminding of Rambus the orphan.

 

[MattZN]

QD > 1 actually isn't rare. Any sequential file access (say, from linux or windows or MacOs or whatever) is actually issuing 4-8 read-ahead I/Os in parallel. So the operating system is, in fact, going to be running the device at QD4 or higher with standard programs, IF those programs try to read the file quickly enough to keep the I/O's pending.

-Matt

 

[MattZN]

For the consumer space, people would be better served with something like the Samsung 951 or 960 EVO or 960 PRO. These devices have a read latency of around 60-70uS and very high sequential bandwidths at QD1 (in excess of 1 GByte/sec) due to the controller having a sequential optimization in its firmware. Beyond that, the Intel numbers are certainly going to be better (at 10x the price!) in the random-read case, but both Intel and the Samsung NVMe devices approach similar performance caps at higher queue depths. So for the consumer it makes *ZERO* sense to buy an expensive XPoint device.

For the enterprise space this first XPoint device has two really serious problems. The first is that chip density isn't nearly high enough. The second is that durability is also not nearly high enough. Intel was originally promising durabilities in excess of 1 million write cycles per cell. They are releasing a device with durabilities in the 30,000 writes per cell range. That just isn't good enough to use as a ram caching layer. Until durabilities improve, I think there will be only modest demand for this particular device in the commercial space. The use case is very specialized. Even most cloud services won't have any use for XPoint.

The lower latency is just not that big an issue in the NVMe form factor... 10uS is still a long time vs dram. For the sub-1uS latency they expect to have in a DDR form factor they need the higher durabilities for it make sense, and the DDR form factor still has serious issues with cpu stalls that a more conventional paging scheme does not have. So in an enterprise environment with a DDR form factor, we not only need far higher durabilities, we also need far more memory channels and far more cpu threads (maybe go to 4-way hyperthreading) to maintain cpu performance.

So Intel might have something here, but the technology needs significant improvement over what they are releasing with this device and their cpus are not well-matched to the increased stalls.

-Matt

 

[palladin9479]

Something like that is WAY outside the range for normal consumer use, it's something designed for database servers or other environments with ridiculous I/O requirements. You want to dump your web content or other high traffic database onto it, or use it as a cache source if your database happens to run into the terabytes in size. Depending on how the stack is architectured this would be placed on the DB server itself or on the front end web proxy server as a huge ass cache. Could even placed it in an ESXi box as a form of hyper fast local storage for VM's, though it would break several other technologies.

 

[nitrium]

QD > 1 actually isn't rare. Any sequential file access (say, from linux or windows or MacOs or whatever) is actually issuing 4-8 read-ahead I/Os in parallel. So the operating system is, in fact, going to be running the device at QD4 or higher with standard programs, IF those programs try to read the file quickly enough to keep the I/O's pending.-Matt

http://www.tomshardware.com/reviews/ssd-gaming-performance,2991-5.html
http://www.tomshardware.com/reviews/ssd-gaming-performance,2991-8.html
http://www.tomshardware.com/reviews/ssd-gaming-performance,2991-11.html
Looks like QDs of 1 to me.

Here is a far more comprehensive report: http://www.xtremesystems.org/forums/showthread.php?260956-hIOmon-SSD-Performance-Monitor-Understanding-desktop-usage-patterns
QDs > 1 really are very rare for consumer workloads. Even under "Heavy Multitasking, he averages a QD of just 1.057.

 

[jarchambeault]

Great review. Very informative and thanks for the info. Is there maybe a typo on the read ave latency / read 99.99% percentile graphs on the Optane QD256? I see 0.44055ms average read latency, but 0.1012ms as the 99.99th percentile 4k read latency, which I read as a max latency. Seems unlikely the average would be 4x larger than the maximum.

- John

 

[PaulAlcorn]

Great review. Very informative and thanks for the info. Is there maybe a typo on the read ave latency / read 99.99% percentile graphs on the Optane QD256? I see 0.44055ms average read latency, but 0.1012ms as the 99.99th percentile 4k read latency, which I read as a max latency. Seems unlikely the average would be 4x larger than the maximum.

- John

Good catch. That was a conversion error on my part (microsecond to millisecond). Those were listed as too low, whereas upon double checking all the values I found that I listed the 70/30 QD percentile measurements too high. Both are updated, thanks for the headsup!