Intel and Micron to Cease Joint Development of 3D XPoint Next Year

[LTC8K6]

Looks like 3D Xpoint may be failing to get any traction...

https://www.tomshardware.com/news/intel-micron-3d-xpoint-imft,37461.html

 

[Cerb]

Do they really expect something that just came out, for all intents and purposes, to get great traction? On the server side, for ones like in the DIMM form factors, software will lag behind the hardware for awhile. Intel wants to confusedly market the product, price the product high, and then after many delays, they are surprised that sales aren't amazing. Plus, the best uses for said product require software or hardware modifications that could take a year or more to either be sufficiently validated for wide use in server farms, or for products to reach the market (expensive DIMMs just replacing PCIe SSDs in your servers doesn't make a lot of sense, unless it's part of an upgrade to even denser form factors, or you use them a bit differently to make better use of them being on a closer interface).

(...) thus incurring under-utilization charges that impacted Micron's bottom line

I think that is much more telling, along with the fact that the changes to their deal include NAND. They want to make and sell their own, and probably without Intel's margins.

 

[whm1974]

Sigh, and 3D Xpoint had such promise. I think that the technology might have taken off had Intel just made SSDs with it instead trying to get everyone to use DIMMs.

 

[mxnerd]

It doesn't mean both companies are abandoning the technology, it just means that they will continue to develop the technology and sell their own products

 

[William Gaatjes]

I think Intel just want to make server oriented products mainly and micron wants to sell serverproducts and customer products.
Maybe Intel and Micron even have a deal that Micron is not allowed to produce dimms similar in technology that are compatible with the new xeons that support 3dxpoint.
With 3dxpoint, i am sure Intel has oncpu memory encryption/decryption as well like AMD EPYC does. Otherwise, someone could just shut down the server, exchange the 3dxpoint module and have all data to be examined while relaxed at home. Of course, restarting the server before leaving.

 

[IntelUser2000]

Intel wants to confusedly market the product, price the product high,

An analyst tweeted saying the costs of making 3D XPoint may mean even the very high $/GB Optane like the 800P may not be making any money for them. Something like $2-2.5/GB.

You could say the server P4800X part is subsidizing the consumer parts.

 

[PliotronX]

I think Intel just want to make server oriented products mainly and micron wants to sell serverproducts and customer products.
Maybe Intel and Micron even have a deal that Micron is not allowed to produce dimms similar in technology that are compatible with the new xeons that support 3dxpoint.
With 3dxpoint, i am sure Intel has oncpu memory encryption/decryption as well like AMD EPYC does. Otherwise, someone could just shut down the server, exchange the 3dxpoint module and have all data to be examined while relaxed at home. Of course, restarting the server before leaving.

That makes sense because the consumer market is pretty cutthroat! It was difficult enough to convince other users to upgrade to SATA SSDs when theyre comparing 120-480GB for more in cost than 1TB-2TB and comparing benchmarks with Xpoint is almost always demoralizing until you use one as its greatest strength is in low queue depths where the majority of workloads take place.

 

[IntelUser2000]

Even now, HDDs outsell SSDs.

 

[Cerb]

An analyst tweeted saying the costs of making 3D XPoint may mean even the very high $/GB Optane like the 800P may not be making any money for them. Something like $2-2.5/GB.

You could say the server P4800X part is subsidizing the consumer parts.

If the costs really are in that ballpark, the retail market doesn't seem like the place to be, period, and sticking mainly to workstation and server markets makes more sense. It's hard to deny the random performance, but it's just not that much better than regular SSDs for most of us, as regular SSDs are plenty fast enough for most of us, and Optane as an old SRT cache replacement has such limited appeal, these days.

In drive arrays for databases or multimedia content, or in extra-low-latency memory bus operation in servers, where they can result in denser total systems, or more performance per sqft with otherwise similar systems, I can see it being worth it (though it may still take a few years for software and hardware configurations to make the best use of it), but for consumer parts, I don't see prices much above $1/GB being worth it.

 

[IntelUser2000]

IMO, the Optane Memory branded devices have their place.

They do say the low capacity parts are there because its using low yielding parts. I'm assuming just like CPUs have different max speed grades, Optane have a grade, and that grade determines reliability. That explains the Optane Memory and the 800P parts only having 6 DWPD equivalent while P4800X is 10x that at 60DWPD(soon anyway).

In terms of volume the lowest is going to end up being the SSDs. It's too expensive in absolute cost for a caching device, while the performance is hobbled by the PCIe interface.

Ultimately RAM is where it'll show its full strength. This will determine whether it has a long future or not.

 

[cbn]

1.) For the low grade Optane (ie, the non-RAM grade 3DXPoint) I am looking forward to Intel allowing it to be used with a RAID array.

Even with 2.5" HDD this could be interesting. Say 4 x 2.5" HDD* (in RAID-5) + Optane NVMe (caching small files) vs. SATA SSD or NAND NVMe SSD.

*If these were 7200 rpm (eg, 2.5" Seagate Barracuda Pro) the Sequential Read of 4 drives in RAID-5 should be equal to that of a single SATA SSD, but with with higher 4K QD1 read (for files cached by Optane).

SIDE NOTE: Romex Primocache does allow Optane to cache RAID arrays even with older systems. See (as an example) results here and here)

2.) For the low grade Optane (ie, the non-RAM grade 3DXPoint) I am wondering if Intel will eventually allow different tuning Options for use with a SATA SSD? For example, perhaps instead of write back we could choose write though (since the high capacity SATA SSDs have fairly strong 4K write on their own. This would increase reliability and allow a greater amount of space on the Optane for 4K Read)

(Eg, 1TB MX500 has 4K QD1 write of 164.2 MB/s vs. Optane 32GB 4K QD1 write of 207.8 MB/s.)

3.) Same idea as #2 goes for low grade Optane (ie, the non-RAM grade 3DXPoint) caching 3D QLC NVMe.

P.S. Would be interesting to see a mobile workstation that combines 3 x M.2 2242 NVMe 3D QLC NAND (parity protected RAID array) cached by a 1 x M.2 2242 Optane using write though. (Compact, (relatively) low power, high performance and reliable)

 

[cbn]

Thinking more about low grade (ie, the non-RAM grade) Optane I am also wondering if we might see it used as MLC? This specifically when it used by the Optane System Acceleration Software for write through rather than write back?

Reason: Write through puts less wear on the Optane compared to write back.

 

[IntelUser2000]

I don't think elaborate wear levelling is necessary on the NVMe parts.

The 16/32GB Optane Memory has 185TBW rating, which is similar to the endurance ratings 256GB driver from the top vendors like 960/970 from Samsung, 760 from Intel.

 

[Cerb]

I don't think elaborate wear levelling is necessary on the NVMe parts.

The 16/32GB Optane Memory has 185TBW rating, which is similar to the endurance ratings 256GB driver from the top vendors like 960/970 from Samsung, 760 from Intel.

That's only ~10k writes, though, at 1x (also, I couldn't find any official write cycle ratings), which would be in the ballpark of commodity MLC as wear leveling became necessary. But, they might just be treating them cautiously, since predicted lifetimes aren't always 100% accurate, and it's fairly new tech.

 

[cbn]

4K QD1 write of some midrange capacity SATA SSDs:

https://www.anandtech.com/show/12263/the-crucial-mx500-500gb-review/5

The 500GB MX500 is down about 45 MB/s compared to the 1TB MX500.........but that is still pretty fast......and probably in the ballpark of a 16GB Optane benchmarked in the same system.

So based on that I would imagine write through mode will be coming to Optane system acceleration software in the near future.

And if write through mode comes, how soon is the next step MLC Optane for NVMe? Would MLC NVMe Optane still have enough 4K QD1 read to make it worthy as cache for an SSD? (I think if it is not true with 1st Gen Optane then it will probably be true with 2nd Gen (or later) Optane)

 

[IntelUser2000]

That's only ~10k writes, though, at 1x (also, I couldn't find any official write cycle ratings), which would be in the ballpark of commodity MLC as wear leveling became necessary.

No, the per cell write is likely much higher.

For example for the 960 Evo 512GB, its rated at 200TB, which equals to 400 write cycles. The 760p isn't much better at 560 write cycles for the 512GB. But TLC cells are rated higher than that, several times in fact! The X25-M with 50nm MLC NAND also ends up at about 500 cycles. Optane Memory does 10K+ with the same metric. Manufacturers counteracted declining endurance ratings by increasingly sophisticated controllers and bigger DRAM buffers.

To be fair, the ratio with Optane Memory probably isn't as high as the NAND drives, but there likely exist a significant gap. The Optane Memory devices have a teeny tiny controller chip to save on costs and power. It probably has a rudimentary wear levelling system and function to convert byte-level access of the media to NVMe compatible 4K block size output.

 

[cbn]

Sigh, and 3D Xpoint had such promise. I think that the technology might have taken off had Intel just made SSDs with it instead trying to get everyone to use DIMMs.

The DIMMs will have a lot of promise besides cache, but one thing to remember is that future NVMe SSDs will get much faster as the PCIe interface increases in bandwidth.

For example, with PCIe 5.0 x 4 (not as far off as most people would think) Sequential Read could scale to as high as 14 GB/s (assuming the controller is up the task). That is a lot of sequential .....but 4K QD1 read will most likely stay around 60 MB/s (on average) for 3D TLC. That makes for a pretty huge gap between NAND's Sequential performance and the 4K QD1 Read.

So having DIMMs selectively cache the smallest files allows the performance to be more level across various file size types. (Assuming enough Optane is available on the DIMM to cover the needs of the smallest files and perhaps some small medium size ones as well).

P.S. Speaking of Sequential Read I am very interested to see how this company's NAND will scale at relatively low capacities.

On paper, the high I/O performance would allow SSD vendors to make low-capacity SSDs with limited NAND channels without today's performance penalty, offsetting the low parallelism with high transfer rates.