3D Xpoint Memory (Intel Optane)

There's a really good article about this at Anandtech but I'm afraid it's time to rein in expectations. This isn't going to replace either your DRAM or your SSD any time soon.

http://www.anandtech.com/show/9470/...000x-higher-performance-endurance-than-nand/1

Key Points:

  • The 10x density is of DRAM rather than NAND so the density compared to NAND probably isn't that great. In fact according to Anandtechs best estimates it's a bit lower than planar NAND. They are also expecting significantly higher per GB cost so it sounds like the dream of an ultra fast multi TB SSD at consumer level pricing is dead for now.
  • Latency is higher than DRAM and while not clear on what the bandwidth will be, it's probably going to be lower than high speed DDR4 and certain to be lower than HBM. So a replacement for DRAM is also off the cards for high end gamers. Also the DRAM modules are initially only being released for Xeons and the timescales for that are unclear.
  • It's being pitched by Intel and Micron as an additional hierarchical step between DRAM and NAND, so for a consumer grade PC it's going to add to your systems cost and complexity while the benefits are likely to be not hugely noticeable. If you can afford enough capacity to completely replace your SSD though, then it should be pretty nice - albeit hugely expensive.
 
Did the announcement contain any claim about power, especially compared to DRAM for the same bandwidth? I'm cautiously excited about this memory being usable in mobile devices if the power makes it worthwhile. If so, it has the potentially to completely change the system architecture and topology of SoCs as we know them today.
 
It usually takes many years to bring a completely new technology to market. If even semiconductor scientists are debating what might be lurking in there, there's a good chance that Samsung etc. are clueless about it as well (despite their tendency of stealing core technologies from competitors, see TSMC.)
In that case, I don't think we'll see a direct competitors anytime soon.

As for prime usage: it's probably most optimal for data center applications. Think memcached instances, where speed is important, but DDR4 speed not so much, yet SSD is too slow.
I think that's where the DIMM form factor will end up.
 
Yes, as the architect of raised floor storage and compute infrastructure, that's exactly where I see this landing. I have two all-flash SAN arrays today that are quite performant, yet we still have two very specific use cases where we've moved to onboard PCI-E flash for the unique hardware boxes. I see this being useful as a tier between SAN-based SSD over 40Ge / 16GFC and the machine itself.
 
I would love a TB of this hung off the processor bus. Sod messing about with filesystems and mmap().
 
If even semiconductor scientists are debating what might be lurking in there, there's a good chance that Samsung etc. are clueless about it as well (despite their tendency of stealing core technologies from competitors, see TSMC.)
I'm not suggesting they have the same tech but there have been a lot of similarish non volatile techs like MRAM & phase change in the pipeline for years.
Some of them have even been in limited production.
Intel going big with their tech may finally get some of those other techs to market properly.


Though, the reported speed boost for the Optane SSD sounds like not really all that much faster than existing SSDs, maybe within range of a new generation controller chip?
 
The fact that the other technologies have been in the pipeline for years is to me more of an argument that they won't be ready anytime soon.

Intel and Micron pulled an Apple here: they announced something with mass production soon, instead of promising something that ultimately never materializes.
 
I already had a post about the tech over at https://forum.beyond3d.com/posts/1867741/ :yes:

Actual timeline SSD versions next year is pretty exciting stuff.
I figure a decent chunk of this stuff as stacked die on-package is where things get really exciting though.

Exciting certainly. I wouldn't expect consumer grade version for another 3-5 years at least, however. I'm expecting the first products in 2016 to be significantly more expensive than their enterprise SSD counterparts (which have significantly higher cost/GB than consumer drives).

Regards,
SB
 
Exciting certainly. I wouldn't expect consumer grade version for another 3-5 years at least, however. I'm expecting the first products in 2016 to be significantly more expensive than their enterprise SSD counterparts (which have significantly higher cost/GB than consumer drives).

Regards,
SB

I'm pretty sure the announcement included SSD's for the consumer market in 2016. Although they will no doubt be a lot more expensive than 'normal' SSD's.
 
Yeah just had a look again and they mention making it available in both 2.5" and M.2 form factors. Rob Crooke is reported to have said it'll be available for data centers and notebooks (I'm betting Apple have already committed to buying a large amount).

So perhaps it'll only have 1.5x - 2.0x the price premium over the current most expensive consumer SSDs.

Intel's prototype is 5-7x the speed of Intel's current fastest NAND based PCIE SSD.

Regards,
SB
 
I can understand why this is coming to market first in SSD form, but I'm more excited by the possibility of non-volatile DRAM-like memory that the underlying technology should make possible.

From what I heard this will actually be coming in the form of a Dimm initially with a SSD form factor to follow, and we should see products late next year or more likely 2017 sometime. The confidence was pretty high there.

Also the subject of wear is an interesting one. I have heard from a storage vendor that in the last 5 years they have seen no SSD's failed due to wear, which is pretty impressive as most applications running on those are DB/VDI/Etc with a fairly high write count. From what I have seen even at current NAND levels of wear endurance you need to write terabytes a day for years to have a failure.
 
There's a really good article about this at Anandtech but I'm afraid it's time to rein in expectations. This isn't going to replace either your DRAM or your SSD any time soon.

http://www.anandtech.com/show/9470/...000x-higher-performance-endurance-than-nand/1

Key Points:

  • The 10x density is of DRAM rather than NAND so the density compared to NAND probably isn't that great. In fact according to Anandtechs best estimates it's a bit lower than planar NAND. They are also expecting significantly higher per GB cost so it sounds like the dream of an ultra fast multi TB SSD at consumer level pricing is dead for now.
  • Latency is higher than DRAM and while not clear on what the bandwidth will be, it's probably going to be lower than high speed DDR4 and certain to be lower than HBM. So a replacement for DRAM is also off the cards for high end gamers. Also the DRAM modules are initially only being released for Xeons and the timescales for that are unclear.
  • It's being pitched by Intel and Micron as an additional hierarchical step between DRAM and NAND, so for a consumer grade PC it's going to add to your systems cost and complexity while the benefits are likely to be not hugely noticeable. If you can afford enough capacity to completely replace your SSD though, then it should be pretty nice - albeit hugely expensive.

I believe the xpoint memory is in the 200-300 nano second access time range, so about 1 order of magnitude slower than dram but several orders quicker than nand. With that said I wonder if that could be a game card DDR replacement? How latency sensitive are graphics/physics engines in that respect?
 
I believe the xpoint memory is in the 200-300 nano second access time range, so about 1 order of magnitude slower than dram but several orders quicker than nand. With that said I wonder if that could be a game card DDR replacement? How latency sensitive are graphics/physics engines in that respect?
GPUs are more tolerant to access latencies, but current architectures are designed for DRAM-range of timings and noting scales for free either. An order of magnitude latency stretch could force a complete resizing of the buffers, pipelines and register files to mitigate the effect and that could distort the fine balance between compute density and efficiency. Personally, I don't see the need for non-volatile memory on a discrete SKU. The IGP solutions are another topic of their own.
 
According to the linked article, the latency mentioned in the presentation seemed to match read latency, not write. Even if it is faster than NAND and "only" an order of magnitude slower than DRAM, it may still have an asymmetry in read versus write that could make it significantly worse.

The endurance question can come up again for high-bandwidth devices like a GPU, since a big one can read/write terabytes in seconds, and the bandwidth/capacity ratio is such that a single device is going to pull tens-hundreds GB/s, unlike an SSD that has 8-10 devices on the far side of a single bus.
 
The endurance question can come up again for high-bandwidth devices like a GPU
Looks like this memory will still be used primarily for long-term storage, only hooked up differently to the system for vastly higher performance than traditional storage subsystems.

Using it directly as CPU/GPU random access memory or cache seems extremely fail, as you could wear it out in days at most in a write-intensive application even if you have a very big pool of memory to work from...
 
Using it directly as CPU/GPU random access memory or cache seems extremely fail, as you could wear it out in days at most in a write-intensive application even if you have a very big pool of memory to work from...
And, compared to other memory technologies, slow and expensive.


Cheers
 
It could be used for datacenters with high read and extremely low write load situations. In that case, it could be used to significantly reduce the power consumption for a large data center versus traditional DRAM which needs to be constantly powered. In a single user system that power savings would be relatively negligible. But with hundreds/thousands of blades/racks it could be significant enough to consider. Especially as it has trickle down effects to how much heat is generated and thus how much additional power must be used to cool the data center.

That's assuming it's at least fast enough to service many read requests in a sufficiently low amount of time.

Regards,
SB
 
Back
Top