Discussion Optane Client product current and future

[nosirrahx]

Looks like Intel plans to up the capacity of the U.2 Version of the 905P to 960GB and 1.5TB and upto 765GB for the M.2 version.

Source

I wonder what will happen to the prices? Things are still in the '$ is no option' class for the high capacity models.

I am still looking for an 8th gen laptop with a 22110 NVMe M.2 port so I can test out the new M.2 Opane drives coming next month.

 

[IntelUser2000]

Nice find @Dayman1225

They have both U.2 and AIC in 1.5TB capacities: https://ark.intel.com/products/codename/67236/Mansion-Beach

At least it should be selling.

Interesting thing about the M.2 version. 765GB version means the package now has 128GB(not 128Gb), up from the maximum of 64GB for the 4 stack version. Maybe its an 8 stack? Or, or, 256Gb XPoint die achieved by 4 layers?

This also has implications for Optane DC PM. They should be able to have a 1TB DIMM stick! That means using Samsung's 256GB DIMMs, a 2P system having 15TB of memory.

 

[Dayman1225]

his also has implications for Optane DC PM. They should be able to have a 1TB DIMM stick!

I suspect they will wait for 2nd Gen to update the DIMMs (Barlow Pass) potentially enabling even higher densities.

 

[Dayman1225]

Intel SSD Road Map from Computer Base

Notice the lack of 96L and XPoint Gen 2, I assume those will arrive in H2'19 then. Weird how TweakTown has such an early IMFT 96L SSD sample then? Perhaps they will go DC first.

 

[cbn]

Nice find @Dayman1225

Interesting thing about the M.2 version. 765GB version means the package now has 128GB(not 128Gb), up from the maximum of 64GB for the 4 stack version. Maybe its an 8 stack?

Yep, must be 8 stack because 2nd Gen doesn't come until sometime in 2019.

https://newsroom.intel.com/news-rel...e-update-3d-xpoint-joint-development-program/

The companies have agreed to complete joint development for the second generation of 3D XPoint technology, which is expected to occur in the first half of 2019. Technology development beyond the second generation of 3D XPoint technology will be pursued independently by the two companies in order to optimize the technology for their respective product and business needs.

 

[cbn]

This also has implications for Optane DC PM. They should be able to have a 1TB DIMM stick! That means using Samsung's 256GB DIMMs, a 2P system having 15TB of memory.

I suspect they will wait for 2nd Gen to update the DIMMs (Barlow Pass) potentially enabling even higher densities.

If they hold back on 1TB DIMMs (using eight 1st Gen Optane dies per package) I wonder if it is because of limitations in Cascade Lake for addressing very large amounts of memory?

Or maybe we see some 1TB DIMM for dense motherboards like the one below from Lenovo Thinksystem SD650........ or motherboards (with more Optane DIMM slots) using Cooper Lake-SP.

 

[Dayman1225]

Pcper appears to confirm that its 8 stacked dies.

Throughout every Intel SSD we have ever laid our hands on for review, we've never seen *any* product (NAND or 3D XPoint) stack more than 4 dies per package. Had Intel stuck with that limit here, we would only have a maximum raw media capacity of 896GB. This is a 1.5TB SSD, so the only possible answer here is that we apparently have the first 8-die-per-package SSD to come out of Intel.

 

[IntelUser2000]

Pcper appears to confirm that its 8 stacked dies.

Interesting because the AIC version only need the 4 stacked package to reach 1.5TB. Now we know why the U.2 version was limited in capacity.

I suspect they will wait for 2nd Gen to update the DIMMs (Barlow Pass) potentially enabling even higher densities.

Hmm. 2x the capacity enabled by 8 stack package, and 2x package with 2nd Gen XPoint for 2TB/DIMM Barlow Pass?

 

[cbn]

Hmm. 2x the capacity enabled by 8 stack package, and 2x package with 2nd Gen XPoint for 2TB/DIMM Barlow Pass?

Then if 16nm rather than 20nm? Another 2x?

4TB DIMM?

 

[IntelUser2000]

Then if 16nm rather than 20nm? Another 2x?

4TB DIMM?

You'll need something like 14nm to reach 2x the density.

By the way, Intel now has a specification page for most of the 905P models, the 100GB P4801X and Optane Memory M10 series.

https://ark.intel.com/#@SolidStateDrives

M10: Performance and power usage goes down as capacity goes down. Performance of the 16GB version is nearly identical to the 16GB Optane Memory. Also code-named Stony Beach. The low active and idle power explains the decent battery life achieved in systems. Still a feat as its not a trivial thing in Windows.

M10-
16GB: 900MB/s read, 150MB/s write, 2W active, 8mW L1.2(same for all 3 capacities)
32GB: 1200MB/s read, 290MB/s write, 2.5W active
64GB: 1450MB/s read, 640MB/s write, 3.25W active

https://www.intel.com/content/www/u...es/enthusiast-ssds/optane-ssd-905p-brief.html

905P M.2 380GB active read 6.5W, write 11.7W, idle, 2.7W

Also updated the main posts.

 

[nosirrahx]

Imagine what the 4KQ1T1 will look like on the first PCIe gen 4 Optane drive. What do you think, 2020 or so?

Intel RAID cards with Optane instead of RAM for cache would also be pretty cool to see.

 

[IntelUser2000]

Imagine what the 4KQ1T1 will look like on the first PCIe gen 4 Optane drive. What do you think, 2020 or so?

4K Random read/write QD1 will NOT improve with PCIe Gen 4. Only sequential bandwidth will. That metric is bound by latency. You'll see much higher figures on the DIMM parts. Sequential bandwidth will likely not go up significantly if the future Optane devices are as power bound as today.

 

[nosirrahx]

4K Random read/write QD1 will NOT improve with PCIe Gen 4. Only sequential bandwidth will. That metric is bound by latency. You'll see much higher figures on the DIMM parts. Sequential bandwidth will likely not go up significantly if the future Optane devices are as power bound as today.

While it is true that as the file size scales down latency overtakes throughput, throughput still will matter at 4K. Sure, it wont be tremendous but between the bandwidth and new controllers the performance bump won't be trivial.

 

[IntelUser2000]

While it is true that as the file size scales down latency overtakes throughput, throughput still will matter at 4K. Sure, it wont be tremendous but between the bandwidth and new controllers the performance bump won't be trivial.

Latency is pretty much throughput at 4K QD1. That's why we see NVMe SSDs with 3.5GB/s sequential being barely better than SATA ones in 4K QD1. The latency reduction is approximately equal to increase in throughput.

For ultra low latency medium like Optane, the primary limitation is PCIe and software overhead. That won't change until its on a DIMM, and it moves to a byte-level access used by the CPU(4K block access is a performance limiter too).

 

[cbn]

Imagine what the 4KQ1T1 will look like on the first PCIe gen 4 Optane drive. What do you think, 2020 or so?

4K Random read/write QD1 will NOT improve with PCIe Gen 4. Only sequential bandwidth will. That metric is bound by latency. You'll see much higher figures on the DIMM parts. Sequential bandwidth will likely not go up significantly if the future Optane devices are as power bound as today.

If that ended up being true then it would be nice if there were additional options in the Intel Optane system acceleration software optimizing for small file size only. In fact, I'm thinking I would probably want this (along with write through rather than write back) for my 58GB Optane SSD if we eventually get the Intel Optane memory application to cache NVMe 3D QLC.

With that noted, smaller lithography should increase performance per watt on future Optane generations.

 

[cbn]

Newegg Price of 32GB Optane has now settled to $47.99 (regular price, not sale price). Likewise 280GB 900p AIC is now $269.99 and the 480GB AIC is now $469.99.

With that noted, my humble opinion is that Optane would be more valuable if it could be cached to 3D QLC NVMe or very low RPM multi-actuator hard drives. Intel, of course, cannot control the availability of very low RPM multi-actuator hard drives......but they do sell 3D QLC NVMe*...and they could make the Intel Optane System acceleration software work better with 3D QLC NVMe by focusing just on small file sizes for caching.

With that noted, I would imagine any optimization that allowed Intel Optane System Acceleration Software to work better with 3D QLC NVMe would also transfer over to very low RPM multi-actuator hard drives. (re: both need the Optane to cache just the small files. Maybe the only difference being 3D QLC NVMe essentially only needs to be "write-through" while the very low RPM multi-actuator hard drive could also benefit from slightly larger file sizes plus "write-back").

*with other companies having 3D QLC NVMe on the way as well.

Intel Optane System Acceleration Software ---> https://www.intel.com/content/www/us/en/support/articles/000024385/memory-and-storage.html

 

[cbn]

it can only work with the os drive

Intel recently updated the Optane memory application so it now works with secondary drives ---> https://www.intel.com/content/www/us/en/support/articles/000027987/memory-and-storage.html

 

[IntelUser2000]

Have to manually choose which drive, not practical at all.

Its not as bad as you say, as its user base is skewed towards enthusiasts, or those much more knowledgeable about computers. Most of the systems being sold to average folks will have it pre-configured with the primary drive.

cbn:

Having it only work with small files makes Optane fit an even smaller niche. Outside of small files, it has the advantage of consistent performance that's not affected by dirty/clean state and % filled with data.

 

[cbn]

cbn:

Having it only work with small files makes Optane fit an even smaller niche. Outside of small files, it has the advantage of consistent performance that's not affected by dirty/clean state and % filled with data.

If comparing a 118 GB Optane SSD to a 120GB NAND based SSD I definitely see your point....but because NAND is so much cheaper the comparison ends up being 118 GB Optane vs. small Optane drive (for cache) + much larger NAND based drive.

So for any given amount of data the Optane cache drive + 3D QLC NVMe drive will not be filled to the same percent capacity as the standalone Optane SSD.

Furthermore, if comparing the Intel 660p SSD to 118GB Optane SSD the Sequential Read is actually faster on the NAND SSD . So caching larger files on the small Optane drive would reduce performance in at least two ways:

1.) Reduced Sequential Read performance.
2.) Reduced capacity on the Optane cache drive for small files.

About the only case I can think for 118GB Optane vs. small Optane cache drive + much larger NAND drive would be if the usage included more small files than the small Optane cache drive could handle.

 

[nosirrahx]

About the only case I can think for 118GB Optane vs. small Optane cache drive + much larger NAND drive would be if the usage included more small files than the small Optane cache drive could handle.

The newer version of Optane cache software is doing something interesting with combining RAM and Optane allowing for better sequential speed.

I have not seen any technical details on this but from the screenshot you can see that the software actually lists the RAM + Optane as "Total Installed System Memory".

The sequential speed is also better than even a 970 Pro which should be impossible.

Does anyone have a 16GB or 32GB setup that can confirm these results? I am wondering if this is some trick specific to the larger Optane modules or if all of them do this now.

 

[IntelUser2000]

Furthermore, if comparing the Intel 660p SSD to 118GB Optane SSD the Sequential Read is actually faster on the NAND SSD . So caching larger files on the small Optane drive would reduce performance in at least two ways:

The Optane products are too high priced to justify just speeding up in one metric. The advantages shown might be even less than today, because performance is much more than just random 4K QD1.

Also,

The 660p is rated at 1.5/1GB/s on the 512GB model for sequential throughput, its random read/write(not 4K) is rated at 90K/220K, which corresponds to 360MB/880MB.

In contrast, out of the 1.45GB/640MB sequential for 800P, the random read/write is rated at 250K/145K, or 1GB/580MB, which is very close to the sequential peak.

So Optane can reach peak rated much more often than NAND drives even in areas outside the random 4K.

https://www.anandtech.com/show/13078/the-intel-ssd-660p-ssd-review-qlc-nand-arrives/6
https://www.anandtech.com/show/12512/the-intel-optane-ssd-800p-review/6

You can see how even with 128KB sequential throughput how the 660p plummets when the drive is full. So even if the 800P is disadvantaged on paper, in reality it'll be able to achieve that all the time while 660p will be massively fluctuating. So it can speed up the 660p in all areas.

But what you really want is bypass all the software and firmware limitations and go straight for using a hardware solution. A single 3D XPoint chip in place of the DRAM buffers used in NAND SSDs and hybrid HDDs.

Even having computer knowledge doesnt make practical sense. It looks like for switching drive all cache woulde be deleted. So if I switch to another drive then have to use the programs in that drive few times to train the memory.

How many of us switch drives? I think you are asking a niche within a niche. Or are you just being a flamebait?

 

[VirtualLarry]

How many of us switch drives? I think you are asking a niche within a niche. Or are you just being a flamebait?

Some of us dual-boot Windows and Linux. We're Enthusiasts, after all, and this is a product aimed at Enthusiasts (arguably), as well as OEMs intended for "factory install and forget" mainstream customers.

 

[IntelUser2000]

Fair enough.

I think he means fully switching a drive. As the cache clearing only occurs when you need to disable it. And you disable it if you want to use the drive again as a standalone or you are upgrading to another drive. If you are dual booting you are still using the same drive. Unless its an external one.

 

[cbn]

https://www.anandtech.com/show/13078/the-intel-ssd-660p-ssd-review-qlc-nand-arrives/6
https://www.anandtech.com/show/12512/the-intel-optane-ssd-800p-review/6

You can see how even with 128KB sequential throughput how the 660p plummets when the drive is full. So even if the 800P is disadvantaged on paper, in reality it'll be able to achieve that all the time while 660p will be massively fluctuating. So it can speed up the 660p in all areas.

This does apply if referencing a $469.99 480GB Optane to a $99.99 512GB 660p, but if going that far up the Optane line then we could also compare it to 118GB 800p caching a 512GB 760p drive.or a 58GB 800p caching a 1TB 760p.

At both capacities (512GB and 1TB) the 760p is 700+ MB/s faster in Sequential Read than the 480GB Optane 900p and doesn't suffer the same slow downs AFAIK as the 660p.

Downside is IOPS. 480GB Optane wins in this department, but I do wonder how important this is for a typical Windows client?

 

[IntelUser2000]

At both capacities (512GB and 1TB) the 760p is 700+ MB/s faster in Sequential Read than the 480GB Optane 900p and doesn't suffer the same slow downs AFAIK as the 660p.

You mean, its 700MB/s faster than the 58GB/118GB 800P, because 900P is faster than 760p even in Anand's sequential benchmarks. The 3200MB/s sequential speed really does not show until really high up in queue depth. Also, while its 700MB/s faster in sequential burst, sustained sequential read shows 900MB/s lead for the 800P over 760p when the drive is dirty.

ALL NAND SSDs suffer from the slowdown FYI: https://www.tomshardware.com/reviews/intel-optane-ssd-905p,5600-2.html
https://www.tweaktown.com/reviews/8626/intel-optane-ssd-905p-960gb-aic-nvme-pcie-review/index6.html

Look on the sequential steady state, and sustained sequential write. Even the higher end drives like the 960 Pro(which is better than the 760p), if you delete a single large file, you can overwhelm the DRAM cache and cause stutter for a brief moment. Of course, that's exceedingly rare.

So Optane drive is only really slower in sequential operations where the drive is relatively empty and its pretty clean, say after a TRIM operation and in high queue depth. All that is rendered quite moot as they are barely faster in loading operations. So if they do what you suggest, they have to optimize for an extremely rare scenario that might be complicated and create a lot of overhead - burst sequential. That's why a clean solution is best for performance.

I did make it clear I'm not a big fan of the Optane SSDs. The future is caching, and DIMMs. DIMMs even on the consumer variant will have much higher sequentials than any SSD, so. NAND SSDs go in between caching and DIMMs rather perfectly.