New memory from Micron/Intel

[Khato]

However I guess they did not use SSDs (that benefit from its built-in controller logic etc), but some other separate NAND chip when testing. The details on that are scarce.

Correct, one of the slides does compare 3D XPoint against NAND across a PCIe NVMe interface with a claimed 10x performance increase - http://www.legitreviews.com/3d-xpoi...nd-micron-to-be-1000x-faster-than-nand_169795 Regardless, I'd expect one of the primary benefits of 3D XPoint versus NAND is that it can be treated like DRAM - no need for fancy controllers to work around the inherent weaknesses of NAND.

As that implies, the primary usage case for this is bye bye NAND for any high performance storage initially... and for everything eventually. Initial pricing is likely going to be high until more capacity is brought online. The question for long-term viability is comparative cost per wafer to 3D NAND - the initial product looks to be a ~217mm^2 die size for 128 Gb... by comparison Samsung's 128 Gb 24 layer MLC is supposedly 133mm^2 and Intel/Micron's 256 Gb 32 layer MLC is around 175mm^2. (Keep in mind that capacity for MLC is number of storage elements times 2.) aka, it's already higher density than 3D NAND with only two layers.

 

[Enigmoid]

Looks good. Now what Intel needs is a fast interface on its chips capable of utilizing it.

Probably needs a lot of work and some sort of controller/interface will have to be added on die.

 

[dark zero]

What Intel need is to revamp the whole Memory System. DDR is starting to slow down the PC and HBM is a temporary solution. It's time for new tech.

 

[Fjodor2001]

If XPoint actually can be made big enough and at a price competitive with SSDs, I could see it replacing SSDs as storage.

It's interesting that Intel envisions that the XPoint memory will be using the DDR4 slots as interface though. Since they say it's not as fast as DDR4 RAM, we might end up with two separate DDR4 banks. One with regular RAM for fast speed, and one for XPoint memory with somewhat slower (but still fast) non-volatile high capacity storage.

 

[Idontcare]

Cursory assessment suggests Intel and Micron have successfully developed a commercially viable memristor-based memory.

edit: no mention of power-consumption metrics or relative comparisons in any way as far as I can see, did I miss it somewhere?

 

[Khato]

edit: no mention of power-consumption metrics or relative comparisons in any way as far as I can see, did I miss it somewhere?

I haven't seen any mention of such either. Which leaves the question of whether it's merely on par with the alternatives and hence not worth including or worse. Would be rather surprising if it's worse, but certainly possible.

 

[Fjodor2001]

One could say the same about HDs vs SSDs.

Sure, but is the market ready for yet another separate memory/storage type in parallel to the existing ones?

We've already got DDR RAM, SSD, HDD, and soon HBM. I think it would be simpler to introduce XPoint memory in the market if it can replace one of the existing memory/storage devices. Likely also cheaper for the consumer, since having several separate devices usually gets more expensive.

 

[Idontcare]

I haven't seen any mention of such either. Which leaves the question of whether it's merely on par with the alternatives and hence not worth including or worse. Would be rather surprising if it's worse, but certainly possible.

First rule in marketing is that you never mention the downside of your own product, only mention the downside of the competing product.

If they have yet to even mention or reference power consumption relative to the other memory types for which they did make latency/performance comparisons then we are safe (justified) to conclude that the power consumption per bit/etc is worse.

The only question I have is "how much worse is it?" Denser than dram, but at what power cost? Faster than NAND, but at what power cost? etc...

 

[ShintaiDK]

Sure, but is the market ready for yet another separate memory/storage type in parallel to the existing ones?

We've already got DDR RAM, SSD, HDD, and soon HBM. I think it would be simpler to introduce XPoint memory in the market if it can replace one of the existing memory/storage devices. Likely also cheaper for the consumer, since having several separate devices usually gets more expensive.

HD usage in 2020 wont be common for example.

 

[Fjodor2001]

HD usage in 2020 wont be common for example.

By then ideally I'd like to see only two memory/storage types in a typical computer; e.g. HBM for really fast GPU and CPU usage, and XPoint/SSDs or similar for non-volatile bulk storage.

 

[ShintaiDK]

By then ideally I'd like to see only two memory/storage types in a typical computer; e.g. HBM for really fast GPU and CPU usage, and XPoint/SSDs or similar for non-volatile bulk storage.

With Intel you will more likely see HMC.

There are 2 cases.
Regular user = HMC memory + XPoint storage.
Server = HMC cache + DDR4 memory + XPoint storage.

Of course there are mixes with legacy tech as well until full crossover points. And DDR4 may dissapear from servers at some future point.

 

[Fjodor2001]

I doubt HDDs will be gone completely even by 2020 though. But most likely they will primarily be used by server centers and similar, and perhaps in homes for bulk storage (e.g. for storing UHD Blu-Ray Rips). After all HAMR is expected to bring 100 TB HDDs by 2025:

Not sure if solid state storage will be available at realistic prices at those sizes by then.

 

[ShintaiDK]

And we still got tapes as well. Doesnt change anything.

 

[Fjodor2001]

And we still got tapes as well. Doesnt change anything.

Yeah, but not very common for regular storage, even in server centers. More or less only for backups. HDDs on the other hand will likely be used for much more than backups in 2020, although it will be much less common for home usage.

 

[kimmel]

If they have yet to even mention or reference power consumption relative to the other memory types for which they did make latency/performance comparisons then we are safe (justified) to conclude that the power consumption per bit/etc is worse.

Being non-volatile we can assume you don't need to spend power to refresh it like DRAM, but yes read power and write power would be interesting numbers to have. They were predictably light on real technical specifics.

 

[MisterMac]

Cursory assessment suggests Intel and Micron have successfully developed a commercially viable memristor-based memory.

edit: no mention of power-consumption metrics or relative comparisons in any way as far as I can see, did I miss it somewhere?

This.

If it's just in between DRAM or a few bit points less - this is a F U to any sort of stacked anything.
...and a revival for the IGP in the future.

1 Pool would of fast storage would make any cpu the last 10 years from intel seem lightning fast to the everyday user.

Not to mention all those DBs and webfarms that would eat this up.

 

[RaynorWolfcastle]

I just found this thread as I had posted something very similar in the memory and storage forums.

I can definitely see something like some near-future platform that would have 8-16 GB of on-package HMC, 128-512 GB of XPoint and TBs of spinning rust or SSD depending on how much you're willing to spend as a consumer.

Looking at it this way, your "RAM" would be faster by an order of magnitude, your "fast persistent storage" would increase by several orders of magnitude (2 or 3 based on the slides?). This would make for a massive improvement in user experience.

 

[Khato]

For those interested, a replay of the announcement webcast is available which includes the Q&A - http://www.intel.com/content/www/us...echnology/intel-micron-3d-xpoint-webcast.html

Few points I found interesting from watching it. First, a bit before 36 minutes in the context of scaling there's a mention that they believe it could evolve to be multi-bit. Second, with respect to cost they vaguely put it between NAND and DRAM, with the important point that it supposedly has no effect on their 3D NAND roadmap. (Implies that manufacturing cost is high enough at current that it's not worthwhile to transition all their fabs to 3D XPoint.) Third, on a question regarding the power and heat of the technology they tip-toed around the question by talking about reduced system power due to instant power off/on and the lack of DRAM refresh power. Lastly, the final question asked what read performance was like under the assumption that the 1000x applied to writes - apparently the improvement for reads is comparable to that for writes.

 

[krumme]

I am on the hype. This is some of the best news the last years.
Its fantastic they can actually already move this tech commercially now for select segments. This early to when this tech was demonstrated from HP the first time in was it 2008. Wow.
And the prospects of multibit in cross points is a freakshow with density and opportunities out of this world.
This is Intel at its best. Lets hope the best.

 

[Phynaz]

Cursory assessment suggests Intel and Micron have successfully developed a commercially viable memristor-based memory.

Holy crap, you're right.

Intel just nuked HP.

 

[tenks]

Ok now I'm really confused. This is different than the HMC on knights landing?

Also, I thought apache pass on purely was the same thing as HMC on knights landing.

Both are compared to better than DDR4 and both are Intel/Micron developed. ?(

 

[witeken]

First rule in marketing is that you never mention the downside of your own product, only mention the downside of the competing product.

If they have yet to even mention or reference power consumption relative to the other memory types for which they did make latency/performance comparisons then we are safe (justified) to conclude that the power consumption per bit/etc is worse.

The only question I have is "how much worse is it?" Denser than dram, but at what power cost? Faster than NAND, but at what power cost? etc...

I was thinking exactly the same.

In the Q&A there was a question about power consumption. The webcast seems to be offline so I can't watch it, but I think they said it's better than RAM because of NVM.

 

[witeken]

And the prospects of multibit in cross points is a freakshow with density and opportunities out of this world.
This is Intel at its best. Lets hope the best.

Yeah, scaling possibilities are great. Go MLC and you have 256Gbit, same as Intel's breakthrough cost 3D. Go to 4 layers and double again. Same for 14nm, etc.

 

[dealcorn]

As a plain vanilla Flash NAND replacement, XPoint lacks a tout worthy energy efficiency benefit. In the, as yet, not fully announced role as a DRAM replacement, Xpoint has big energy efficiency benefits. You already know that SSD's are more efficient than DRAM. When Intel adds a small DRAM cache to enable XPoint to function as a DRAM replacement, XPoint retains the efficiency advantage of SSD's because it is non volatile and may be powered off. The Breezy article is all over this and the cited patents are the foundation for his conclusion that

The Intel/Micron PCMS tech facilitates such power efficiency that they will be able to offer a no-compromise PC in a smart phone form factor.

There are standards concerning how quickly SSD's recover from DevSleep to fully active. Eventually, a deep dive article may contrast the SSD DevSleep recovery vs the XPoint comparable. In due course, testing will compare and contrast the performance of cached XPoint as a DRAM replacement vs actual DRAM. Much is unknown.

 

[R0H1T]

As a plain vanilla Flash NAND replacement, XPoint lacks a tout worthy energy efficiency benefit. In the, as yet, not fully announced role as a DRAM replacement, Xpoint has big energy efficiency benefits. You already know that SSD's are more efficient than DRAM. When Intel adds a small DRAM cache to enable XPoint to function as a DRAM replacement, XPoint retains the efficiency advantage of SSD's because it is non volatile and may be powered off. The Breezy article is all over this and the cited patents are the foundation for his conclusion that

The Intel/Micron PCMS tech facilitates such power efficiency that they will be able to offer a no-compromise PC in a smart phone form factor.

There are standards concerning how quickly SSD's recover from DevSleep to fully active. Eventually, a deep dive article may contrast the SSD DevSleep recovery vs the XPoint comparable. In due course, testing will compare and contrast the performance of cached XPoint as a DRAM replacement vs actual DRAM. Much is unknown.

Yeah four billion dollars & counting, btw that was just for last year & this time they've hid it in the balance sheet. That's such a vague statement that it's hard to know whether anyone's taking the (desktop) PC market seriously at all :hmm: