New memory from Micron/Intel

[Fjodor2001]

A press release is hype? I don't think you understand the word.

They are feeding the news media with the info, in order to hype it. Quite obvious, and I don't see how you can fail to see that.

 

[Khato]

There first product will be 128 gigs 2d not 3d so yes this XPoint is still in the future.

The initial product according to the PR is 128 Gb across 2 memory layers, so technically still 3D. It's just not a huge number of layers like is being done with 3D NAND. Which I'm guessing is mostly due to the fact that it's not necessary given that Samsung's 24 layer 3D NAND is also a 128 Gb die. (aka, if Intel released a 16 layer version with the same capacity per layer they'd have 128 GB on a single die.)

 

[Tuna-Fish]

I said the news articles said it's 1000x faster. Normally that would mean it has 1000 times higher bandwidth unless else is specified.

This news directly came from a slide that used the 1000x faster to refer to having 1/1000th of the latency of NAND flash.

Assuming you can accept large enough transfer sizes and some parallel accesses, bandwidth of memory is driven by the interface, not the memory technology itself. HBM, GDDR5, DDR3, and even plain old SDR DRAM use very similar memory arrays running at quite similar speeds. The massive bandwidth difference is not gained by having faster memory (mostly, there has been some advancement but it's been shockingly slow by semiconductor standards), just by accessing more memory cells in parallel.

Similarly, the bandwidth of this device has almost nothing to do with the memory technology itself and almost everything to do with the interface it attaches to. If it attaches to a DDR4 socket, well, it will have bandwidth similar to DDR4.

 

[ShintaiDK]

The initial product according to the PR is 128 Gb across 2 memory layers, so technically still 3D. It's just not a huge number of layers like is being done with 3D NAND. Which I'm guessing is mostly due to the fact that it's not necessary given that Samsung's 24 layer 3D NAND is also a 128 Gb die. (aka, if Intel released a 16 layer version with the same capacity per layer they'd have 128 GB on a single die.)

Considering they use DDR4 DIMM types. I think we can estimate capacities to be 128-256GByte per DIMM with 8-9 chips per side. But it could also be more, thats seen previously.

 

[Fjodor2001]

This news directly came from a slide that used the 1000x faster to refer to having 1/1000th of the latency of NAND flash.

Assuming you can accept large enough transfer sizes and some parallel accesses, bandwidth of memory is driven by the interface, not the memory technology itself. HBM, GDDR5, DDR3, and even plain old SDR DRAM use very similar memory arrays running at quite similar speeds. The massive bandwidth difference is not gained by having faster memory (mostly, there has been some advancement but it's been shockingly slow by semiconductor standards), just by accessing more memory cells in parallel.

Similarly, the bandwidth of this device has almost nothing to do with the memory technology itself and almost everything to do with the interface it attaches to. If it attaches to a DDR4 socket, well, it will have bandwidth similar to DDR4.

So you're saying that it's actually 1000x times faster, and can have a bit width comparable to other memory technologies? That would allow for 1000x bandwidth as well.

Sure, if it is limited by DDR4 specs, it won't be any faster than what that spec allows. But if what you're saying is true, the memory technology itself actually has the potential of 1000x higher bandwidth, if attached to an interface that allows it. Sounds too good to be true...

 

[jpiniero]

I'm sure it will be only a server technology for some time. The power savings alone will make many server customers VERY interested. It's not as fast as DRAM but it sounds like it would be close enough to make the tradeoff worth it in plenty of server scenarios.

 

[Tuna-Fish]

So you're saying that it's actually 1000x times faster, and can have a bit width comparable to other memory technologies? That would allow for 1000x bandwidth as well.

Sure, if it is limited by DDR4 specs, it won't be any faster than what that spec allows.

1000x times faster than NAND on the workload they tested, which was probably writes, as NAND can at least in theory read a lot faster than it can write. As per their own published marketing material, it's latency is somewhat worse (a full order of magnitude?) than DRAM.

But if what you're saying is true, the memory technology itself actually has the potential of 1000x higher bandwidth, if attached to an interface that allows it. Sounds too good to be true...

What I want to reiterate is that almost all (flash writes being the primary exception, as cell clears have so high energy use that a very high bandwidth device would melt) solid-state memory technologies have the potential of near-limitless bandwidth through parallel access, if the workload calls for it. It's really an issue of interface tech, not the ram itself. The limit is cost and interface power, not the devices holding the bits.

If Intel believes that it is sensible to ship this on a DDR4 DIMM, it is indeed good news about it's expected performance.

 

[dark zero]

Oh boy.... DDR is more likely to being doomed then... So... maybe Canonlake and Zen+ along Volta and Artic Islands sucessor will bring dramatic changes on performance on everything. Interesting...

 

[R0H1T]

Oh boy.... DDR is more likely to being doomed then... So... maybe Canonlake and Zen+ along Volta and Artic Islands sucessor will bring dramatic changes on performance on everything. Interesting...

You really think so? What about HBM/HMC as a natural evolution to DDR(4) as I don't see any PMC or anything similar giving you a bandwidth of a TBps (& way more to come) anytime soon & well the latency will also be an issue.

 

[witeken]

There first product will be 128 gigs 2d not 3d so yes this XPoint is still in the future.
In the video they claim how games are limited because they have to keep going into the system for more info and how this bottleneck can be replaced with XPoint tech.
Now suppose you added XPoint L4 cache on the package.

It's 2 layers, so 3D. And it's 128Gbit like SSDs, not GB.

 

[witeken]

BTW, no coverage from AnandTech? Where are the deep dives we used to get from Anand Shimpi?

 

[Fjodor2001]

1000x times faster than NAND on the workload they tested, which was probably writes, as NAND can at least in theory read a lot faster than it can write.

In reality SSDs which use NAND can write about as fast as they read these days. In fact some can even write faster than they read. See e.g.:

http://www.anandtech.com/show/9451/the-2tb-samsung-850-pro-evo-ssd-review/6

http://www.anandtech.com/show/9451/the-2tb-samsung-850-pro-evo-ssd-review/7

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.

 

[R0H1T]

In reality SSDs which use NAND can write about as fast as they read these days. In fact some can even write faster than they read. See e.g.:

http://www.anandtech.com/show/9451/the-2tb-samsung-850-pro-evo-ssd-review/6

http://www.anandtech.com/show/9451/the-2tb-samsung-850-pro-evo-ssd-review/7

Writes should normally be faster on an SSD anyway cause you don't (necessarily) have to write data contiguously, on the other hand reads are a bit more of an uphill task because they are not always contiguous hence the fall in speeds.

 

[Fjodor2001]

I wonder what the primary use cases for this new memory are?

If it's slower than RAM (as the slides indicate), it cannot replace that. So will it be big enough to replace SSDs? If not that either, I'm not sure where this new memory type fits in...

 

[R0H1T]

I wonder what the primary use cases for this new memory are?

If it's slower than RAM (as the slides indicate), it cannot replace that. So will it be big enough to replace SSDs? If not that either, I'm not sure where this new memory type fits in...

It's just a niche product that'll do good for (large) enterprises & even then only those firms which require PCM or some such headache, cause adding another layer of memory is just that! It won't replace DDR4 anytime soon & it cannot touch HBM/HMC (definitely not second gen either) plus it won't match NAND prices anytime soon, so all there's left to do is to see how much of this niche segment grows & whether it'll be the next big thing in computing or just another fad, like HVD's

I'm not saying that it'll die away, like HVD, but just that it won't (necessarily) be a phenomenon like NAND in recent years.

 

[witeken]

I wonder what the primary use cases for this new memory are?

I'd say it's a very good successor of NAND for high-end applications. I'd also have said mobile, but they're being vague about power consumption, so I fear it's not impressive.

It seems it's not good enough to replace RAM, but we'll see. In any case it's a whole lot cheaper than RAM and it has Moore’s Law on its side.

 

[ShintaiDK]

I wonder what the primary use cases for this new memory are?

If it's slower than RAM (as the slides indicate), it cannot replace that. So will it be big enough to replace SSDs? If not that either, I'm not sure where this new memory type fits in...

Funny this comes from someone who else blindly demands more speed

You basicly ask why use SSDs over HDs.

 

[Fjodor2001]

Funny this comes from someone who else blindly demands more speed

You basicly ask why use SSDs over HDs.

Read my post again. I said that if it's big enough to compete with SSDs, then that's of course a use case. But I get the impression that this new memory type will not be big enough to compete with SSDs (at reasonable price). Got some info indicating otherwise?

 

[NTMBK]

BTW, no coverage from AnandTech? Where are the deep dives we used to get from Anand Shimpi?

In a drawer at Apple.

 

[ShintaiDK]

Read my post again. I said that if it's big enough to compete with SSDs, then that's of course a use case. But I get the impression that this new memory type will not be big enough to compete with SSDs (at reasonable price). Got some info indicating otherwise?

One could say the same about HDs vs SSDs.

Microsoft is already full aboard this and got a new filesystem for it. Specially databases will be night and day difference, even compared to SSDs.

 

[witeken]

Application of this, for gaming: https://www.google.com/patents/US20140198116

From the excellent Seeking Alpha article that predicted this 2 months ago: http://seekingalpha.com/article/3253655-intel-and-micron-the-purple-swan#comments_header

* A GPU with 1 GByte of DRAM augmented with 64 GBytes (unheard of in today's age of 2-4 GB GPUs) of PCMS and discusses the implications of high-fidelity 3D textures with no loading delays between different sections of the in-game "world."

* This is a good patent to read in its entirety because there is much discussion related to managing the tremendous complexity of making this all work to create something that's years ahead of the competition. Even stereoscopic frame buffers are mentioned. Gamers will open their wallets for this technology.

 

[R0H1T]

Application of this, for gaming: https://www.google.com/patents/US20140198116

From the excellent Seeking Alpha article that predicted this 2 months ago: http://seekingalpha.com/article/3253655-intel-and-micron-the-purple-swan#comments_header

* A GPU with 1 GByte of DRAM augmented with 64 GBytes (unheard of in today's age of 2-4 GB GPUs) of PCMS and discusses the implications of high-fidelity 3D textures with no loading delays between different sections of the in-game "world."

* This is a good patent to read in its entirety because there is much discussion related to managing the tremendous complexity of making this all work to create something that's years ahead of the competition. Even stereoscopic frame buffers are mentioned. Gamers will open their wallets for this technology.

Remind us the bandwidth PCM will provide? HBM gen1 is at 512GBps excluding a potential overclock.

A general patent filing tells you as much about its potential as Apple's patent on rectangles, or was it rounded squares
The fact that Intel hasn't included PCM for Kabylake should tell you something about its potential in the consumer arena.

P>S> this needs to be moved to mem & storage asap!

 

[witeken]

FYI, the first application Intel gave in its presentation yesterday was gaming.

And as you can read it's used in conjunction with RAM.

 

[R0H1T]

FYI, the first application Intel gave in its presentation yesterday was gaming.

And as you can read it's used in conjunction with RAM.

You know how everyone reads presentations on this forum don't you OR are Intel presentations more reliable than anyone else out there?

Tell me how will that work, a GB of GDDR5 with 64GB of PCM with the latter possibly replacing NAND at some level? Unless they're gonna price SSD's out of the current (evolving) market I don't see how this will interest most consumers, let alone major GPU makers & their rivals AMD/Nvidia & without them it'll never take off in the GPU space. Also SSD's are not a bottleneck in games, unless everyone wants microsecond level of speedup I don't see any real need for this to get into the space of SSD's (at least for games) & cost will also be a major hurdle. Let's see if corporate adoption rates make it cheaper for consumers, for once I wouldn't mind them paying the early adopter tax

 

[dealcorn]

There may be some confusion about timing. Select data center application value the greater speed, endurance and efficiency of XPoint more highly than anyone else so they likely will pay a premium for all available manufacturing capacity for some time. Because XPoint appears to be both cheaper to manufacture than regular flash NAND and possesses superior performance metrics, XPoint is a strong candidate to replace up to 100% of all flash NAND used in SSD's as additional capacity comes on line. Some have questioned whether Samsung can make it's inferior performing 3D NAND cost competitive. It may be challenging for Samsung to defend market share with a product that performs worse and costs more the manufacture than XPoint. Performance testing and market share data will provide clarification in due course. Till then, this is speculative.

Stephen Breezy's above linked Purple Swan article speaks of an Intel patent with target:

The focus of this approach is on providing performance with a relatively small amount of a relatively higher-speed memory such as DRAM while implementing the bulk of the system memory using significantly cheaper and denser non-volatile random access memory (NVRAM).

Use of XPoint as a generic system RAM replacement will likely turn on announcement of a product that implements this approach. Intel's gaming slide may represent this patent. Gamers typically are willing to pay premium prices for performance. Due to efficiency and big cost benefits (vs DRAM), the natural mainstream XPoint memory product introduction may target ultamobile markets like smart phones. Breezy also make the point that Xpoint provides big benefits to the data center in terms of reduced UPS costs. He may be right. For non enthusiast markets, I see strong appeal for an affordable XPoint on package boot device that will never wear out in typical consumer use. The BoM benefits will be big and almost everyone needs an affordable, good performing boot device.

It takes time for capacity to come on line but ultimately Xpoint targets all flash NAND and multiple big chunks of the DRAM market. Other than that, it is no big deal.