The Intel Optane SSD 900P 280G Review(anandtech.com)
anandtech.com
The Intel Optane SSD 900P 280G Review
https://www.anandtech.com/show/11953/the-intel-optane-ssd-900p-review
48 comments
The niche for this device is high random read performance in workloads where the dataset is too large to fit in RAM, for a significant price premium. Random write and sequential read/write performance is poor value compared to cheaper and higher capacity NVME ssds from Samsung.
The one standout feature apparently is that performance is reliable and not subject to GC stalls or degradation as the drive fills up.
The one standout feature apparently is that performance is reliable and not subject to GC stalls or degradation as the drive fills up.
500k sustained random IOPS from a single device is remarkable. To put that in perspective, 500k PIOPS in AWS would cost you $32,500/month (https://aws.amazon.com/ebs/pricing/).
This drive looks like it would be amazing for database workloads, for example.
This drive looks like it would be amazing for database workloads, for example.
I remember when the rule of thumb for big storage arrays was "100 IOPS per disk." It's still only that with common desktop disks and only 200-ish for enterprise SAS disks. This device provides the equivalent of 2500-5000 disks of IOPS. Latency, particularly write latency, should be better than flash as well.
The price isn't too bad really; a Samsung 960 PRO M.2 1TB is $600 for a little more than twice the storage. It's still early days for 3d xpoint; Intel and Micron will build on this. Exciting stuff.
The price isn't too bad really; a Samsung 960 PRO M.2 1TB is $600 for a little more than twice the storage. It's still early days for 3d xpoint; Intel and Micron will build on this. Exciting stuff.
An EC2 I3 has up to 3.3 million IOPS.
https://aws.amazon.com/about-aws/whats-new/2017/02/now-avail...
https://aws.amazon.com/about-aws/whats-new/2017/02/now-avail...
"Up to". Also we don't know how many underlying SSD devices there are, which is my point. This is a single drive that supports radically more sustained random IOPS than we've ever seen, at a very low price point compared to comparable solutions.
That's with eight 1.9 TB devices. You can scale it down, of course.
Anyway, the point is you do not need to spend anywhere near $32,500/m to get that kind of performance from AWS. (As long as you don't mind using non-EBS storage.) In this case, an i3.16xlarge is $27,856 per YEAR (if paid upfront).
Anyway, the point is you do not need to spend anywhere near $32,500/m to get that kind of performance from AWS. (As long as you don't mind using non-EBS storage.) In this case, an i3.16xlarge is $27,856 per YEAR (if paid upfront).
I like to use my AWS ops to send SMS messages so I can really get the most out of my service providers.
Remind you, transfer throughput improvements from HDD to SDD had limited effect on perceived performance; the reduction in latency was what made SSDs magical (compare normal USB stick to USB sticks with SSD controller). Optane when they manage to get it to original promise and optimize software around it, should yield a similar latency reduction, making them feel much snappier than current SSDs, though throughput-wise they might not be so shiny.
I didn't see in the article anything talking explicitly about latency reduction. I know SSDs have much lower latency than HDDs; could you expand / provide stats on how 3D Xpoint's latency compares to SSD?
I have an Intel i750 NVMe SSD. It has about a 60 microsecond latency.
This device is supposed to be on the order of 10 microsecond.
The only people that would notice that are people running database servers serving hundreds of thousands of users.
This device is supposed to be on the order of 10 microsecond.
The only people that would notice that are people running database servers serving hundreds of thousands of users.
I wonder what the relative production cost is for the 3D Xpoint vs NAND. Given the newer state of the technology it seems like there might be more low hanging fruit in terms of performance/cost improvement.
If Intel can hit a production cost per GB close to standard NAND it seems like they would be able to absolutely dominate the premium consumer/enterprise SSD market.
If Intel can hit a production cost per GB close to standard NAND it seems like they would be able to absolutely dominate the premium consumer/enterprise SSD market.
The production cost is probably pretty high: https://www.anandtech.com/show/11454/techinsights-publishes-...
Does anyone know if there are stats on things like Windows boot times / game load times? I've seen graphs that show boot times comparable between average SSDs and the 960 EVO; is this something an average home user could feel an improvement with?
Booting Windows and loading games is full of large sequential reads using high queue depths. This is something ssd controller can optimize well already.
Average consumers don't need this. Enthusiast gamers don't need this. This is enterprise hardware rebranded for consumers.
Average consumers don't need this. Enthusiast gamers don't need this. This is enterprise hardware rebranded for consumers.
No, it's exactly the opposite.
Enterprise workload with a big queue depth is the realm of SSD.
Xpoint brings the same or much better speed (use case dependent) even to the consumers that need for 99% of the time a queue depth between 1 and 4.
Its super exciting for new storage technologies to be released, but at this time it doesn't seem that much better than Samsung's Evo 960, which can be had for much lower prices.
It is aeons better if they deliver:
* High Read And Write Performance At Small Capacity Points
* High Performance Maintained As The Drive Fills
* Heavy Workloads That Cause Drive Wear And Require High Endurance
Because SSD fail all three:
They require huge blocks to be wiped if you want to write anything, they grind to halt once you fill past 50%, and they eventually degrade.
From what I've seen in charts, they deliver.
* High Read And Write Performance At Small Capacity Points
* High Performance Maintained As The Drive Fills
* Heavy Workloads That Cause Drive Wear And Require High Endurance
Because SSD fail all three:
They require huge blocks to be wiped if you want to write anything, they grind to halt once you fill past 50%, and they eventually degrade.
From what I've seen in charts, they deliver.
"aeons better" is like "made the Kessel Run in less than twelve parsecs"
Do you think "generations better" is bad, too? Aeons are like that but more intense.
I think 'generations better' is also lousy but at least 'generations' as 'stages of technological development of something, over time' is common figurative usage. 'Aeons' isn't, making it much sillier. Intensely sillier, if you prefer.
Aeons are commonly used for revolutionary stages of something. Not necessarily improvement, but definitely giant leaps of progress. It's not just a measure of raw time.
I can think of a retort that is ages and eras better than this!
Agreed. The only thing keeping me from putting these into production is that I can't find any data on reliability.
EVO 960 isn't even in the same league as this. EVO can maybe possibly burst to a double digit percentage of this performance at its best moment, but under long-term mixed load is more than an order of magnitude slower.
There are some graphs in the review that show the 900p to be an order of magnitude faster than the 960, like "Sequential Steady-State Mixed Workload", which is a workload typical for a busy server.
I'm hoping that changes as they improve yields / gain more commercial adoption.
This is "Intel" we're talking about. I'm no expert on the industry, but from what I recall Intel is the king of fabs.
This is "Intel" we're talking about. I'm no expert on the industry, but from what I recall Intel is the king of fabs.
Intel is also king of margins. Considering that NAND is 4x the density of XPoint I wouldn't expect it to get much cheaper.
Well, they've also chased quite a fee geese that haven't provided any golden eggs yet. Like Xeon Phi, which was supposed to be the GPGPU killer for half a decade now. But it's never really been more than temporarily on par with NVidia's offerings.
"Intel's greatest invention... was AMD."
;D
;D
The 900P is also not THAT much faster than the 2+ year old Intel SSD 750 series PCIe cards which cost about half as much per GB of storage.
https://www.intel.com/content/www/us/en/products/memory-stor...
https://www.intel.com/content/www/us/en/products/memory-stor...
I have Intel 750 and that cheap Optane Memory, my Optane Memory is significantly faster in fsync speed when tested with postgresql's pg_test_fsync
Do you seriously believe that Intel 900p is not going to be better than Optane memory in that regard?
Do you seriously believe that Intel 900p is not going to be better than Optane memory in that regard?
Looks like its almost completely worthless for consumers. Real world performance is the same as NVMe drives yet at much higher $/GB
That's... not at all what I'm seeing. Very broadly: at high queue depths and on mostly sequential data, it's competitive with and slightly better than the flash devices. At low queue depths and on random data (pretty much exactly where "consumers" live) it's much, much faster. Like, 4x the scores in s a bunch of those tests.
Now... to be fair consumers applications in the modern cloudworld are generally not storage-limited. So maybe your point is defensible in a specious sense. But as actual performance measurements go, these really are ripping fast drives.
Now... to be fair consumers applications in the modern cloudworld are generally not storage-limited. So maybe your point is defensible in a specious sense. But as actual performance measurements go, these really are ripping fast drives.
> We've never experienced a system that felt as fast and responsive as our workstation with the 900P as the operating system drive, and that’s largely due to the explosive random read performance at low queue depths. At QD1, the 900P is nearly 4x faster than the Samsung 960 EVO 1TB
Are we reading the same article?
Are we reading the same article?
I would expect anybody working with DAWs and sample libraries (extremely low latency random read performance needed, meaning every note playing can be a sample anywhere on your drive) would jump at these, both pros and amateurs.
I've been thinking about moving my Kontakt libraries to a fast SSD. Would you see a difference noticeable between this and Samsung 960 for that purpose?
I'm assuming it depends from the size of your template (i.e. how many samples at a time you're using). For not huge templates you can already run kontakt fully purged off the SSD, but I would assume with this the limiter would be pretty much entirely the CPU... still I don't know of anybody having done any DAW tests with XPoint and large sample libraries (say, Spitfire etc.), would be interesting to see what kind of polyphony a high core CPU with this would be able to achieve...
Some stress testing with older SSD's here -
https://vi-control.net/community/threads/performance-to-be-e...
I use Spitfire stuff, mainly Albion, but the voice requirements these guys have is on another level. The Optane speedup would certainly mean more voices, and perhaps remove the need to have separate SSD's for each large kontakt library?
I use Spitfire stuff, mainly Albion, but the voice requirements these guys have is on another level. The Optane speedup would certainly mean more voices, and perhaps remove the need to have separate SSD's for each large kontakt library?
yeah, vi-control is a really good forum for these discussions :) I am waiting for a post from Chimuelo saying he's tried optane, as he seems to regularly try new hardware. I would think that optane would be amazing for this and would basically make your CPU the limiter as opposed to your storage, on the other hand it makes me wonder if Kontakt can take full advantage of this.
I wish benchmarks included C++, C#, js (gulp? Webpack?) and Java builds so that programmers knew whether an upgrade is worth it.
Price is truly amazing. I expected 280GB to cost around $700. This thing is going into my next PC.
very disappointing compared to the original marketing promises.
I think I need to get one of these.
1. Write Endurance is amazing, you can keep writing 2.8Tbytes per day every day for 5 years, as a comparison, Intel 750 only allow you to write 70GBytes per day. Such high write endurance does raise an interesting question - why still bother paying for those expensive enterprise NAND SSD.
2. Random IO performance is identical to P4800x, what Intel actually took away from P4800x or is 900P just a slightly different product with a slightly revised spec list. ;)
For me, this is the most exciting hardware release in 2017, the price point is just amazing, it is going significantly boost the performance I am going to get from my distributed consensus library which requires fsync for each write.