Discussion Optane Client product current and future

[IntelUser2000]

If the CPU isn't keeping the SSD 100% busy with read commands, then its power draw is not going to be 1.85W any more. The drive will be spending some of its time idle (approximately 32% idle in this example).

You are assuming an ideal world, where everything works perfectly. The scenario you suggested can easily be true for a CPU. However we know that dozens of different clock levels exist even in C0 power state. Whether the CPU decides to go to idle between the bottleneck times or goes to power down is a tradeoff it makes between performance and power consumption. While it may be more efficient to go to lower C states and turn part of the CPU off, it may also decide to run at lower frequency, because transition states are far faster within C0 than between Cx states. Clearly, they felt the need which is why they implemented the 3 PSx states.

 

[IntelUser2000]

I wonder if this one has the same controller as the upcoming Carson Beach?

This hit us out of nowhere. Are we going to see a 900P version of this?

I don't think this is Carson Beach. This is probably a higher end drive meant for enterprise. The controller is same as the P4800X. Mansion Beach refresh maybe?

Also, its longer than regular m.2 drives. It looks like M.2 22110 or something.

https://www.servethehome.com/new-intel-data-center-optane-m-2-ocp-summit-2018/

The drive is longer in dimensions relative to its width compared to the 800P.

 

[cbn]

Yep, it is longer.

And I am assuming it has 12 dies? (four dies in each of the three packages).

So if the 280GB 900P has 21 dies and 280GB (out of 336GB).....this one (If made into a lower end 900P model) would have 160GB (out of 192GB)?

P.S. If it did get released I do think it would be very interesting as a option to the 118GB 800P and the 280GB 900P. (Personally I much rather have a 160GB 900P (with the better controller) than a 118GB 800P for desktop....though I do acknowledge that it would not work in most (all?) laptops as I believe they can only take 2280. (In fact, it looks like even these laptops can't take 22110)

 

[Dayman1225]

https://www.anandtech.com/show/12562/intel-previews-optane-enterprise-m2-ssd

AnandTech have more details, same controller as P4800x/900P but shrunk down, lower performance expected due to heat and power constraints. Uses 7 quad die packages to have upto 375GB of storage so far, to be named P4801x.

 

[cbn]

https://www.anandtech.com/show/12562/intel-previews-optane-enterprise-m2-ssd

AnandTech have more details, same controller as P4800x/900P but shrunk down, lower performance expected due to heat and power constraints. Uses 7 quad die packages to have upto 375GB of storage so far, to be named P4801x.

7 packages for 7 channels. Nice.

I still wonder if they might release one with 3 packages? (Eg, We saw a version of Optane Memory with one less package than the other) This mainly because the Sequential performance would be so much better than the 118GB 800p.

My estimate for a 160GB (192GB) "900p" with three packages (4 dies each):

Sequential Read: ~2500 MB/s*
Sequential Write: ~1800 MB/s* (re: each die is good for about ~150 MB/s, so 12 dies x 150 MB/s= 1800 MB/s)

* With enough cooling.

^^^^ That is a big jump in Seq. performance for a mere ~33% increase in capacity.

 

[IntelUser2000]

Sequential Read: ~2500 MB/s*
Sequential Write: ~1800 MB/s* (re: each die is good for about ~150 MB/s, so 12 dies x 150 MB/s= 1800 MB/s)

I don't think it'll work that way. The full 900P gets 2000MB/s with 7 channels, regardless of capacity. The full drive can have 28 chips. I'm not sure why the 16GB Optane Memory only gets 145MB/s.

Intel has a datacenter NAND SSD that uses the M.22110 format. It's the DC P4501. The M.2 ones are discontinued though. It has 8.25W load and <3W idle in its highest capacity configuration.

Eg, We saw a version of Optane Memory with one less package than the other

I have the 16GB Optane Memory, and has only 1 package.

Searching on Optane M10 shows a result on a foreign site with ETA for April 4th. That's the same week for 300-series chipset launch. Ok, I can't rely on ETA as 800P said Feb 14th. At the moment though seems like M10 would be a good candidate for "Next Gen Optane Memory". And it would be more than a coincidence if that happened to be Carson Beach.

 

[nosirrahx]

I made an interesting discovery regarding the 900P. I asked Intel if it is possible to use the 900P as Optane cache (since its larger and faster than their initial offerings) and they said no. Turns out, it is completely supported.

 

[cbn]

I made an interesting discovery regarding the 900P. I asked Intel if it is possible to use the 900P as Optane cache (since its larger and faster than their initial offerings) and they said no. Turns out, it is completely supported.

Very interesting.

So the entire 900P can be used for cache? Not just 64GB of the 900P?

 

[nosirrahx]

Very interesting!

So the entire 900P can be used for cache? Not just 64GB of the 900P?

Its hard to tell for 100% sure (since Intel claims this wont work) but it seems so. The Optane setup takes forever and does seem to be caching the entire OS and apps and then goes on to zero the unused portion.

After reboot the system is insanely fast, noticeably faster than the SATA drive was. I posted about this over at guru3d and left a link of this system booting (post #5):

https://forums.guru3d.com/threads/tried-something-interesting-with-a-u-2-900p-and-it-worked.419655/

BTW, on a related note, I also discovered a bug in Optane that affects the re-use of an Optane drive.

If you use the Optane app to disable the cache and then use that same Optane drive on new install it will appear to work but the Optane app will show that Optane memory is disabled and the optimize task will stay permanently blank.

Intel knows about this, I reported it in a thread where another user had the same issue and I was able to replicate it.

 

[IntelUser2000]

You cached a 4 Terabyte SSD with an Optane drive? Wow, love to be that user lol. By the way, the boot speed is insane.

You were saying on that thread that you tested the Optane Memory as well? The QD1 results should be identical, and in some cases even better for the Optane Memory than the 900P. Did it feel much faster using the 900P as a cache as opposed to Optane Memory? That's assuming you tried both setups on the same system.

It does cache the entire drive. The point of the Optane Memory application is to make it transparent as possible, so the two drives become one.

 

[nosirrahx]

You cached a 4 Terabyte SSD with an Optane drive? Wow, love to be that user lol. By the way, the boot speed is insane.

You were saying on that thread that you tested the Optane Memory as well? The QD1 results should be identical, and in some cases even better for the Optane Memory than the 900P. Did it feel much faster using the 900P as a cache as opposed to Optane Memory? That's assuming you tried both setups on the same system.

It does cache the entire drive. The point of the Optane Memory application is to make it transparent as possible, so the two drives become one.

The 900P as Optane cache was slightly slower than the 900P on its own as a standalone SSD but nothing you would ever notice. Compared to the SATA SSD on the other hand it is dramatically faster.

All of this started with a goal of 4TB and speed with a single drive letter. I first checked into the 16/32GB drives but the gap between them and a typical NVMe drive was not super impressive and the cache capacity was also not what I waned. The 900P crushes the 16/32GB drives in every way unfortunately including the price.

 

[nosirrahx]

FYI, this all started here:

https://communities.intel.com/thread/122265

 

[cbn]

Intel H310 doesn't support Optane memory:

https://www.pcworld.com/article/326...-lake-8th-gen-core-desktop-cpus-chipsets.html

That is very surprising.

 

[cbn]

https://www.anandtech.com/show/1260...-mobile-iris-plus-desktop-chipsets-and-vpro/5

Ultimately, Intel is still marketing Optane, its high-cost R&D product, with low capacities at low cost systems with relatively little margin. While it might seem like a noble goal, to bring extra caching performance down to the lower cost segment, it could very easily be done with SATA or PCIe M.2 drives using regular NAND flash. A lot of users would like to see high-capacity, high-endurance Optane drives moving more into the mainstream, instead of more attempts at funneling in a product like Optane into caching.

In order to help out the situation I think it would be great if Intel would explore the idea of altering the Optane memory software so that the entire drive doesn't need to be used as cache. This way the remaining portion could be used for page file. (My guess is that this would work especially well when a hard drive is the primary drive).

P.S. If writes on the Optane drive need to be a certain level (for good pagefile expansion) perhaps the minimum size for optane cache could be 16GB. This way a person would need a 32GB Optane if they wanted to use drive for page file in addition to cache.

NOTE: More testing needed to see if 32GB Optane (on an Intel system) has enough write speed for page file expansion during browsing.

EDIT: I need to look into how the existing Intel Optane software handles the page file.

 

[VirtualLarry]

Intel H310 doesn't support Optane memory

Intel, IMHO, has Optane positioned exatly wrong. It's the lower-end systems, budget ones that ship with HDDs that need Optane caching. Higher-end rigs will just ship with a larger primary SSD, thus making Optane virtually useless in its current market positioning.

 

[cbn]

Intel, IMHO, has Optane positioned exatly wrong. It's the lower-end systems, budget ones that ship with HDDs that need Optane caching. Higher-end rigs will just ship with a larger primary SSD, thus making Optane virtually useless in its current market positioning.

Agreed.

So right now my hope for the low end rests almost entirely on a 20th Anniversary Celeron that I am guessing would have to be BGA.

 

[IntelUser2000]

Optane Memory now supported on mobile, and on secondary storage.

Intel, IMHO, has Optane positioned exactly wrong.

Ehh. They should have made it to the H310. But B360 is fine. The bigger deal in terms of pricing is the CPU. Enabling Pentium Gold support for example.

But this is due to the way Intel thinks. They see it as a way to upsell their platforms. I disagree on this aspect, and they could be better served by having Optane sold in higher volumes by enabling it on lower CPUs. It would have enabled faster proliferation of the technology. Then, they could do the SKU play.

While it might seem like a noble goal, to bring extra caching performance down to the lower cost segment, it could very easily be done with SATA or PCIe M.2 drives using regular NAND flash.

No, I disagree with this quote. Optane is a much better cache drive. A cache drive that needs to do garbage collection is not a good cache drive. With NAND flash cashing, you have two scenarios for plummeting performance, one is running into the limits of HDD when the usage scenario is demanding enough to outstrip the cache capacity. Two, is when it runs into the limits of DRAM cache on the SSD itself. NAND running out of DRAM cache is sometimes worse than an HDD-only system. NAND further slows down when the drive capacity is close to full. Optane has none of those problems.

The Optane Memory software also pre-loads some OS files so you get some benefit right away. Based on some reviews, the benefits are apparent even on applications with first load.

 

[cbn]

IMO In order to help proliferate Optane on lower end SKUs the technology needs to be seen as a low end form of system memory rather than a high end SSD.

Maybe a 20th Anniversary Celeron (or Celerons) is the time to do that?

P.S. Interestingly the Celeron 300A (released 1998) was innovative in that its L2 cache moved to the processor itself....so I got to wondering if maybe a 20th Anniversary could do something similar and move the optane from the M.2 to the processor package?

https://www.anandtech.com/show/174

By decreasing the amount of L2 cache from 512KB on the Pentium II, to 128KB and placing it on the processor's die itself (meaning that the L2 cache is an actual part of the processor, and not part of a separate package as it was with the Pentium Pro and as it is with the Pentium II Xeon) Intel could produce a low-cost yet high performance solution which would carry on the Celeron name for quite some time. What did Intel call this new processor? The Pentium II Celeron 300A.

With this noted, I would still be more than happy with a 20th Anniversary Celeron capable of using the standard M.2 version of Optane. More than happy, indeed!

 

[Insert_Nickname]

Intel, IMHO, has Optane positioned exatly wrong. It's the lower-end systems, budget ones that ship with HDDs that need Optane caching. Higher-end rigs will just ship with a larger primary SSD, thus making Optane virtually useless in its current market positioning.

If you don't mind the hassle standalone Optane drives make excellent boot drives. They're more then capable of matching high-end 1TB NAND SSDs in low queue depths. At a far lower price.

 

[IntelUser2000]

P.S. Interestingly the Celeron 300A (released 1998) was innovative in that its L2 cache moved to the processor itself....so I got to wondering if maybe a 20th Anniversary could do something similar and move the optane from the M.2 to the processor package?

Moving the cache to the CPU package was done for performance reasons. There'd be no gain doing it for Optane Memory. At least not until they get the DDR interface version ready.

Just having it supported would be nice.

 

[cbn]

Moving the cache to the CPU package was done for performance reasons. There'd be no gain doing it for Optane Memory. At least not until they get the DDR interface version ready.

Wouldn't the latency be a lot lower for on package Optane compared to M.2 interface optane? Maybe even lower latency than Optane on DIMMs?

P.S. The implementation doesn't have to be used just on Celeron. They could also have it with more dies stacked on package for the Core models as well. (In fact, with Intel moving towards co-branding Optane with Core by using the Core+ designation think the increased integration would be of high quality.)

 

[IntelUser2000]

Wouldn't the latency be a lot lower for on package Optane compared to M.2 interface optane? Maybe even lower latency than Optane on DIMMs?

That's why I said until the DDR interface version is ready.

The latency by being on package will only be lower if its on the DDR interface. If its on-package but NVMe, it'll bring no benefits.

Then I am not sure if there will be a noticeable latency benefit by moving it on package. Sure, it'll be physically closer, but by then, the latency of the media may not justify moving it on package. There may be a small benefit, but 3D XPoint on a DIMM is expected to have several times greater latency compared to a DRAM.

If future 3D XPoint iterations are good enough to replace DIMMs in vast majority of the scenarios, we may see Intel moving the controller on-die rather than having something like an FPGA on each of the modules so it can lead to lower cost.

 

[cbn]

That's why I said until the DDR interface version is ready.

Well, we do know they will have that tech ready for server later this year.

http://www.tomshardware.com/news/intel-xd-xpoint-dimm-lenovo-thinksystem,36573.html

Intel's Apache Pass DIMMs make that a reality in Lenovo's new ThinkSystem SD650 servers, which are the first servers to support the new devices. These new DIMMs will feature 3D XPoint memory addressed as normal system memory in a RAM slot, but they aren't on the market yet.

Then I am not sure if there will be a noticeable latency benefit by moving it on package. Sure, it'll be physically closer, but by then, the latency of the media may not justify moving it on package. There may be a small benefit, but 3D XPoint on a DIMM is expected to have several times greater latency compared to a DRAM.

That is an interesting point you bring up (makes sense too), but something else to consider is that by having the controller and Optane on package it can be cooled by one larger heatsink and it would be more compact. (I am actually thinking of 4.5W to 15W processors the most when I think of this kind of implementation, but I would think the 28W processors could work too.)

The other thing to consider is timing vs. NVDIMM-P? (I like the idea of NVDIMM-P....a lot!, but isn't that going to take some time to get going even if the specification gets approved this year?)

P.S. Would be great to see atom come with some integrated Optane as well.

 

[cbn]

While I do think it does seem a bit odd that Intel would put Optane on package and not system DRAM remember Intel does not make DRAM (except eDRAM). Furthermore including Optane on package lets them do some product segmentation.

With this noted I do think we will eventually see System DRAM on package in the form of higher density HBM type memory. So maybe HBM type memory on die with the Optane on package?

(This increased amount of variation in chips beyond what we normally see possible via EMIB?)

EDIT: Just thinking maybe if the capacity for HBM type memory gets high enough there could even be stacks comprising of DRAM on the bottom with Optane stacked on top. This not for the Optane latency improvement, but for the Sequential read/write power efficiency improvement* and possibly the enhanced durability of the Optane).

*Optane that would normally run very slow at reduced wattage could regain speed via the increased parallelism of the HBM technology.

 

[IntelUser2000]

With this noted I do think we will eventually see System DRAM on package in the form of higher density HBM type memory. So maybe HBM type memory on die with the Optane on package?

You see, the reason Intel doesn't make DRAM is because once they exited the DRAM business, their fabs were tuned for logic. Logic transistors are built for speed, not so much density. If you look at their eDRAM, its built on their process, and several times larger than DRAM dies for the same capacity. It also allowed Intel to build eDRAM without having a fab dedicated to one. So what you are suggesting is unlikely. HBM chips are also built on a memory focused process not logic.

With HBM on-package, Intel will have to source from other manufacturers. The future memory hierarchy is:

DRAM moves to HBM on package
SSD moves to Optane DIMMs
Rest filled by high capacity NAND SSDs

So rather than needing a DRAM DIMM per Intel DIMM, HBM on package could serve the same duty, so your DIMM slots can be filled entirely by Intel DIMMs.

I am actually thinking of 4.5W

As of right now, the packages for the 4.5W Core chips and Intel small core ones are quite filled, and have no room for the extra die.