Discussion Optane Client product current and future

[IntelUser2000]

A dram-less NVMe NAND BGA SSD?

Haha, sounds like a dream. I hope they still ditch the software based solution for a firmware one so its not prone to issues and its compatible with all platforms that use NVMe.

 

[cbn]

Here is the Intel Ark link for M15:

https://ark.intel.com/content/www/u...es/189225/intel-optane-memory-m15-series.html

I am very interested to find out the 4K QD1 write for these.

 

[cbn]

The 58GB 800p is faster than the 32GB M15 in Sequential write (640 MB/s vs. 550 MB/s) while also being slightly lower power (3.75W vs. 3.8W). This while also being on the old controller.

Still 550 MB/s for the 32GB is super impressive. This and 300 MB/s for the 16GB M15 is the first time I have seen proof that Gen 1 Optane dies can over ~160 MB/s Sequential write per die. (IntelUser2000 was right that the Sequential write per die could be increased with better power management coupled to enough TDP.)

 

[cbn]

Haha, sounds like a dream. I hope they still ditch the software based solution for a firmware one so its not prone to issues and its compatible with all platforms that use NVMe.

Maybe it will just be a NAND package then.

And Intel uses Foveros chip stacking to couple a NAND controller to the Optane controller.

 

[Billy Tallis]

I am starting to wonder if the 815P is a myth.

It's not. There were too many authentic leaks. The 815P was definitely on the roadmap last year. Since the H10 and M15 were both delayed significantly, it would be no surprise to see the 815P launch in the next few months. If we make it past FMS without hearing anything more, then it will be reasonable to suspect that their plans for the 815P have changed by more than just the release date slipping.

 

[IntelUser2000]

Still 550 MB/s for the 32GB is super impressive. This and 300 MB/s for the 16GB M15 is the first time I have seen proof that Gen 1 Optane dies can over ~160 MB/s Sequential write per die.

It can do better. The 256GB version of Optane DC Persistent memory gets 3.0GB/s write bandwidth with 8 chips. The 128GB does a bit lower at 2.5GB/s or so, but its limited to 15W rather than 18W maximum for 256GB and 512GB version. Interestingly, the 256GB is the fastest out of the three.

This while also being on the old controller.

Yea, it seems like extra read is contributing to this as well. Same with NAND, the PCIe 4.0 SSDs are needing beefy heatsinks.

 

[cbn]

So now that Optane M15 Read latency is down to 6 us I wonder what the 4K QD1 Read will be on Icelake vs. non-Icelake systems?

As a comparison here is the read latency of other Consumer Optane SSDs.

900p/905p: 10 us
Original optane memory: 8 us
M10/800p: 7us
H10: 7us

EDIT: The following Intel document lists both M10 and M15 with read latency 6.75 us---> https://www.intel.com/content/www/u...ptane-memory/optane-memory-m10-m15-brief.html


(Would also be interesting to compare 4K QD1 Read on Zen 2 (which has hardware fix for Spectre)).

 

[IntelUser2000]

Their documents are little off sometimes. The individual page for the drives show the 64GB M15 being at 6us, while the 64GB M10 is at 7us. I guess it could be an average?

It could also explain why early release documents were showing slightly higher sequentials than actual. The ARK pages also seem to take a somewhat conservative approach to specifications because the measured sequentials are often little bit higher.

Interesting they got the latency down despite some articles saying they chose PCIe x2 being lower latency.

 

[IntelUser2000]

Next gen Optane Memory listed as a feature for Intel's 10th Gen U mobile platform code-named "Icelake".

 

[cbn]

Laptops at Best Buy with Optane H10:

https://www.bestbuy.com/site/searchpage.jsp?_dyncharset=UTF-8&id=pcat17071&iht=y&keys=keys&ks=960&list=n&sc=Global&st=intel optane h10&type=page&usc=All Categories

(7 of the 10 listing use the 512GB and 3 of the 10 listings use the 1TB.)

Looking at the HP Envy x360 it has an i7, 8GB and 512GB Optane 10 for $999.99 ( Two comparison points are the HP Envy x360 with i7, 8GB and 1TB HDD for $899.99 and a HP Envy x360 with i5, 8GB and 256GB (non-Optane) SSD for $849.99.)

Overall, not bad.


P.S. Here are Storage configurations for Envy x360 from the HP website:


https://store.hp.com/us/en/Configur...Id=&catEntryId=3074457345619203822&quantity=1

https://store.hp.com/us/en/Configur...Id=&catEntryId=3074457345619013337&quantity=1

 

[arandomguy]

I feel the problem for Optane on the consumer side is the large fall in NAND prices over the last year and half. This has greatly extended the price differential between NAND drives and Optane. As such I wonder how much this effects Optane (or 3D xpoint as a whole) product plans for the consumer market.

 

[cbn]

I feel the problem for Optane on the consumer side is the large fall in NAND prices over the last year and half. This has greatly extended the price differential between NAND drives and Optane. As such I wonder how much this effects Optane (or 3D xpoint as a whole) product plans for the consumer market.

Hypothetically by having Optane in front of NAND it would allow the NAND to be developed more for capacity rather than endurance.

Think Optane + 16nm 3D QLC vs. 20nm 3D QLC.

(Ideally the Optane would also write directly to QLC rather first to SLC NAND. This would reduce power consumption and overhead.)

 

[nosirrahx]

I feel the problem for Optane on the consumer side is the large fall in NAND prices over the last year and half. This has greatly extended the price differential between NAND drives and Optane. As such I wonder how much this effects Optane (or 3D xpoint as a whole) product plans for the consumer market.

People are willing to spend crazy $ for top performing products so if Optane was a simple drop in solution it would find a home in the enthusiast market.

High capacity SSDs with low cost NAND paired with high speed Optane would be a cool solution if the actual software lived on the drive itself allowing them to simply appear as a single NVMe SSD to the system.

The hybrid drives offer nothing that I want. I will continue to pair a low cost SSD with an 800p (and later 815p) cache drive and continue to ignore everything Intel wants me to do with their product.

 

[arandomguy]

People are willing to spend crazy $ for top performing products so if Optane was a simple drop in solution it would find a home in the enthusiast market.

High capacity SSDs with low cost NAND paired with high speed Optane would be a cool solution if the actual software lived on the drive itself allowing them to simply appear as a single NVMe SSD to the system.

The hybrid drives offer nothing that I want. I will continue to pair a low cost SSD with an 800p (and later 815p) cache drive and continue to ignore everything Intel wants me to do with their product.

I'm not really discussing this specific product but the entire plans for Optane/3d xpoint being affected by the large NAND price drop and continued downward pressure. On a tangent I feel this also lowers the attractiveness of QLC NAND as well due to the sudden lower relative savings (but this is separate topic, I was thinking about posting about).

NAND SSD prices have fallen by 50%+ over the last year with still strong continued pressure downwards. Optane product prices (at least retail to consumers) from what I can tell have been basically stagnant. It was just a flash sale but here in Canada Amazon even had the SX8200 1TB (TLC/NVMe) at the equivalent of $95 USD. If TLC NVMe SSDs reach close to $100 and SATAs for sub $100 for Black Friday this year this dramatically changes the cost attractiveness of other options including Optane solutions compared to if they were still $300+.

The consumer market is still going to be sensitive to price. Everyone is going to have cut off points and as the price gap grows more will drop out. My original plans were a Optane/3d xpoint + SATA storage setup. But if for the price of a 128GB optane drive say I can get a 256GB NVMe NAND drive? Optane for sure but what about 512GB? 1TB? 2TB? Eventually at some point more and people are going to be finding the single NVMe more attractive. And we may be approaching the point were even a 2TB NVMe drive will be cheaper than a 128GB optane drive.

While there are some that will pay for marginal improvements at extreme costs they are still the minority. At the end it's question of the groups total spending power in terms of how much attention (resources) they draw from companies in making products.

Enterprise and Optane is of course a different matter. The cost to performance (or gain) calculation is also different there.

 

[nosirrahx]

https://www.intel.com/content/dam/t...es/product-quick-reference-matrix-2019-Q3.pdf

First doc I have seen with model numbers for the next gen Optane products.

SSDPEK1F118GAX1 <- retail 118GB 905P

MEMPEK1F064GAX1 <- retail 64GB M15

 

[IntelUser2000]

High capacity SSDs with low cost NAND paired with high speed Optane would be a cool solution if the actual software lived on the drive itself allowing them to simply appear as a single NVMe SSD to the system.

You still need DRAM on top of that. So Intel needs a controller that can perform decently as a top performing NAND controller plus make it work with Optane.

Enterprise and Optane is of course a different matter. The cost to performance (or gain) calculation is also different there.

It also sucks on Enterprise, IMO. They talk about sales on Optane Memory but not for Optane SSDs.

I have always felt the SSDs are going to be a temporary thing. Too much of the media advantages are lost on the SSD form factor. Optane DC PMs on the other hand, are great.

I feel the problem for Optane on the consumer side is the large fall in NAND prices over the last year and half.

Also they needed to bring the Memory Drive feature to client. Make "Optane Memory" not yet-another-caching solution but act as a pseudo extension to RAM. Then they would eventually evolve the "Optane Memory" being true, Memory, in a DIMM slot.

Based on cbn's experiences, I think using today's Optane Memory with Memory Drive will work. It's surely better than going on the SSD as virtual memory. Send it straight to Optane!

 

[thor23]

Ram's relatively cheap too, I can blockcache(I use primocache but there's other block caching solutions too) my os drive for up to 95% of diskreads with only 8Gb of cache, meaning if I had a spare 8Gb ram I could run my system entirely from ramdisk(I am actually getting another 16Gb ram for this purpose).
For home user this is what optane is competing against unless it can lower it's price to ssd levels.
Main reason I'm getting 16Gb more of ram instead of more optane module is because the price of the 32Gb/58Gb modules isn't cheap enough vs ram.

 

[IntelUser2000]

Main reason I'm getting 16Gb more of ram instead of more optane module is because the price of the 32Gb/58Gb modules isn't cheap enough vs ram.

Persistence makes it a better caching solution(because you don't need to load every reboot), but yes you are right.*

It's still cheaper per GB, but the advantages of greater bit density solutions like Optane(even NAND) only shows in larger capacities. 32GB SSDs are not that cheap, its in the 128/256/512GB capacities where the advantages really show.

So expect Optane to get advantages over RAM in not offering cheaper memory, but by offering greater capacities. Future Optane won't be cheaper, but at the same, or slightly higher cost offer significantly higher capacities.

*Needing to load is the real problem, and where people are looking forward to superfast NV memory like Optane. That's why I'm expecting the DIMM version to be the one that shines.

 

[quikah]

Persistence makes it a better caching solution(because you don't need to load every reboot), but yes you are right.*

It's still cheaper per GB, but the advantages of greater bit density solutions like Optane(even NAND) only shows in larger capacities. 32GB SSDs are not that cheap, its in the 128/256/512GB capacities where the advantages really show.

So expect Optane to get advantages over RAM in not offering cheaper memory, but by offering greater capacities. Future Optane won't be cheaper, but at the same, or slightly higher cost offer significantly higher capacities.

*Needing to load is the real problem, and where people are looking forward to superfast NV memory like Optane. That's why I'm expecting the DIMM version to be the one that shines.

I have been working with Optane DC Persistent Memory. This is speaking from the Enterprise side, I have no idea what Intel plans for desktop regarding this.

You can use it in 2 modes, either memory mode where it acts as main memory or appdirect mode where it acts as storage. Optane memory is slower than DDR4. This is mitigated by using DDR4 memory as a cache in memory mode (generally you use half your slots with DDR4 and the other half with Optane DC PMEM). As soon as you overflow the cache you can halve your performance on very latency sensitive benchmarks because of the higher latencies of Optane memory. Memory mode doesn't really seem that useful to me, sure you can get huge memory sizes for slightly lower prices, but worse performance. Very limited use cases IMO.

Appdirect mode is more useful, but there are downsides. First you are using up your DIMM slots for this, so you are limiting the amount of memory you can have. Second, you need support for it. There are multiple levels of support you can have, the OS can just expose it as a drive, but then you need to go through the various OS layers to get to it which imposes some minor performance penalties. Alternatively you can modify your appliction to be aware of it bypassing the OS. So the application would allocate a certain range of memory which it knows is persistent.

 

[IntelUser2000]

I have been working with Optane DC Persistent Memory. This is speaking from the Enterprise side, I have no idea what Intel plans for desktop regarding this.

That's very cool.

I'm aware of the latency/bandwidth issues. However, when they had the distinction with EP/EX Xeons, the EX Xeons used buffer chips which added significant latency/power/cost.

It won't be useful for all scenarios, but the market is much, much larger than the SSDs.

 

[jpiniero]

Have to wonder if it would be useful to use Optane PM via PCIe 5 CXL instead of DIMMs.

 

[nosirrahx]

I finally got my 4X 905P setup up and running (and bootable). I ran into a snag with my Asus Hyper 16X. It turns out that it has crap power delivery and was corrected in a their V2 card.

https://imgur.com/MCa857u

 

[IntelUser2000]

Have to wonder if it would be useful to use Optane PM via PCIe 5 CXL instead of DIMMs.

That'll result in significant increase in latency similar to Optane SSDs.

I finally got my 4X 905P setup up and running (and bootable). I ran into a snag with my Asus Hyper 16X. It turns out that it has crap power delivery and was corrected in a their V2 card.

The numbers look nice. Do you actually notice a difference in things you do?

 

[thor23]

Sequential speeds are very close to my ramdisk speeds on(I get 13Gbytes/s from my ramdisk cache). Only problem is raid hurts 1q1t 4k performance which isn't that important if it's some multi user cluster, but for a single user the 1q1t 4k is most important.

 

[IntelUser2000]

but for a single user the 1q1t 4k is most important.

It matters for system memory, but there are limits to QD1 performance on disks. At some point, it doesn't do much. The Optane SSDs are quite past that already. The reason likely has to do with the fact that storage is considered "cold", and much as possible is shuffled into RAM.

I think it'll be ridiculously involved to take advantage of even 70MB/s QD1 4K performance.