News Intel celebrates the arrival of MRDIMMs — a plug and play solution for ultrafast memory that offers double the memory bandwidth of standard DRAM

[bit_user]

Micron's official claims:

• ...
• Up to 40% latency improvements compared to RDIMMs**

** Empirical Stream Triad data comparing 128GB MRDIMM 8800MT/s against 128GB RDIMM 6400MT/s.

Well, that's potentially misleading. When we talk about memory latency, most of us mean best-case latency. What they're clearly talking about is closer to worst-case latency, which is the only way you can achieve lower latency purely by means of higher bandwidth.

I've seen the same sort of argument made for using HBM on CPUs, because its best case latency is actually worse than regular DDR memory. However, the argument is that what you really care about (especially on a server CPU under load) is latency under medium and heavy load. By providing much greater bandwidth, HBM is able to keep queues less full and thereby delivers better latency under those conditions.

 

[EzzyB]

Which are they? Are they the Granite Rapids-compatible flavor or the JEDEC flavor?

They are the only flavor.

By working with a single supplier, Nvidia was captive to their pricing and supply constraints. The benefit of a standard is that customers have interoperability with different suppliers, turning the the object into a commodity and providing all the benefits of an open market.


A de facto standard is when something that's not an official standard fills the same role as one (i.e. to enable multiple producers and consumers). So long as Micron was the sole producer of GDDR6X memory, it cannot be considered a de facto standard. Micron's announcement only said that consumers weren't limited to Nvidia - it didn't say they're giving the rights for other DRAM makers to produce competing chips.

Here's where we differ. If you and I worked in business in 2005 and I needed to send you a document I would send you a .doc file. You would be able to read it. It's what everyone used. Sure .pdf files existed and probably even had an official standard (and were created just for that purpose). But that's not what we'd use because .doc was the de facto standard. If you bought a PC game at the same time it has a better the 90% chance of being DirectX game. Sure OpenGL was an official standard, but few made games to meet that standard (some did, of course). Microsoft, because of it's market dominance, set those standards whether we liked them or not.

So, in my view this becomes the de facto standard. 100% of MRDIMMs are now manufactured this way. The lack of a JEDEC rubber stamp only means it's not a JEDEC standard. Perhaps Intel and AMD are so far apart on this that we'll never have a real standard here. I don't begin to understand the technical reasons other than AMD doesn't seem to have a chip yet that can take advantage of these modules. That might be all it is and there is some ridiculous politics going on here. (That actually makes the most sense but it seems disingenuous to say so without actual knowledge of it.)

I don't agree that something can't become a standard without some industry group blessing it. Pretty much everyone involved would prefer it that way, but something has caused JEDEC to drag it's feet here. The obvious answer, at least, is that AMD and/or ARM has some technical issue with it. If you are Intel, you simply can't afford to lose the advantage MRDIMM gives you at a time when you are already losing market share. This is almost certainly how many unofficial standards get started.

 

[bit_user]

I don't agree that something can't become a standard without some industry group blessing it.

But they need to have the rights to the IP. That's where JEDEC plays a role as an IP clearing house. There are lots of patents involved in new memory technologies and nobody wants to sink hundreds of $Millions (or more) into design and manufacturing of a new memory product, only to face litigation that threatens to rob most of their profits from it.

Adhering to JEDEC standards also keeps everyone in the same boat, IP-wise, so that if someone like Rambus comes along and starts litigating, it should hit everyone equally.

something has caused JEDEC to drag it's feet here.

How do you know? Maybe Intel just started working on it sooner and got out ahead of the JEDEC version? Advanced memory technologies are tricky, because you not only have to prove that something is technically possible to do at scale, but also economically viable to implement. That involves a lot of theoretical work, designing & manufacturing test chips on the latest nodes, and lab testing. Maybe iterate that process a few times. Meanwhile, everything has to be codified in a standard, circulated, reviewed, questions & concerns addressed, re-reviewed, and finally approved by the committee. It's a long process.

This is almost certainly how many unofficial standards get started.

CISCO was pretty well-known for that. They would get out ahead of a standard with an implementation, and then try to use their market power to coerce the standard towards matching their implementation as closely as possible. Or so I've been told.

In the RISC-V world, there was a counter example of a Chinese company that created their own ISA vector extensions (maybe even based on some early draft), but the consensus went in a different direction and they got stuck with CPU cores with dead-end vector ISA extensions that nobody wanted program for and toolchains didn't even want to support. So, getting ahead of a standard can definitely backfire.

 

[thestryker]

How do you know? Maybe Intel just started working on it sooner and got out ahead of the JEDEC version?

Given that they contributed the specs to JEDEC this is almost assuredly true.

 

[EzzyB]

But they need to have the rights to the IP. That's where JEDEC plays a role as an IP clearing house. There are lots of patents involved in new memory technologies and nobody wants to sink hundreds of $Millions (or more) into design and manufacturing of a new memory product, only to face litigation that threatens to rob most of their profits from it.

Adhering to JEDEC standards also keeps everyone in the same boat, IP-wise, so that if someone like Rambus comes along and starts litigating, it should hit everyone equally.

I certainly wouldn't argue that standards aren't a good thing, they most certainly are. But, if you don't believe that politics gets involved, well, that's a bit naive. JEDEC, like all standards committees is supported by and funded by a lot of really big corporations with huge financial stakes.

How do you know? Maybe Intel just started working on it sooner and got out ahead of the JEDEC version? Advanced memory technologies are tricky, because you not only have to prove that something is technically possible to do at scale, but also economically viable to implement. That involves a lot of theoretical work, designing & manufacturing test chips on the latest nodes, and lab testing. Maybe iterate that process a few times. Meanwhile, everything has to be codified in a standard, circulated, reviewed, questions & concerns addressed, re-reviewed, and finally approved by the committee. It's a long process.

I don't know. I tried to make that clear. But, both AMD and Intel submitted MRDIMM specs to JEDEC two years ago. Intel has a product right now that can use it. As far as I know AMD and ARM don't. JEDEC hasn't completed the standards work. Why is that? SK Hynix has a MCRDIMM product out. Micron has a MRDIMM product out. These companies are the recognized world leaders in memory. Absolutely no one has more expertise in the designing, manufacturing and testing of memory modules. And these companies have invested millions in those products.

So, I don't know. But, what this duck looks like, and smells like and quacks like is a disagreement among the JEDEC committee members. This is how standards are absolutely not supposed to work, but there are billions at stake and it wouldn't be the first time that some members of such an organization drag there feet until their own companies can catch up. Remember, there is no JEDEC employee here who is some neutral paragon of the holy value of standards. These committees are composed of representatives of the various commercial companies who have a financial stake in the standards. (As an example the chairman of the JC-45 DRAM committee is listed as representing Samsung on the JEDEC web site.)

CISCO was pretty well-known for that. They would get out ahead of a standard with an implementation, and then try to use their market power to coerce the standard towards matching their implementation as closely as possible. Or so I've been told.

In the RISC-V world, there was a counter example of a Chinese company that created their own ISA vector extensions (maybe even based on some early draft), but the consensus went in a different direction and they got stuck with CPU cores with dead-end vector ISA extensions that nobody wanted program for and toolchains didn't even want to support. So, getting ahead of a standard can definitely backfire.

It will be interesting. I can't believe that Micron doesn't have a huge influence here. It's easy to frame this as AMD vs ARM vs Intel, but you already have the largest memory manufacturer in the world producing them. So that's why I think this just ends up being MRDIMM 1.0.

 

[bit_user]

JEDEC, like all standards committees is supported by and funded by a lot of really big corporations with huge financial stakes.

It's an industry consortium, so it basically is its member companies.

JEDEC hasn't completed the standards work.

Maybe, but do we know that, for a fact? It should be completed before there are products on the market. So, the lack of products doesn't imply the standard isn't complete.

SK Hynix has a MCRDIMM product out. Micron has a MRDIMM product out.

I think it primarily requires the multiplexer chip be added to the DIMMs. It would be interesting to know if there's a 3rd party multiplexer and if that's what both are using or if either made their own. The point of MCR/MR is that it's supposed to work with commodity DDR5 DRAM, so it wouldn't necessarily be a big lift for Micron to offer this memory, particularly if they can use an existing multiplexer from a 3rd party.

The key point of uncertainty would be the availability of the IP, but if Intel transferred that to JEDEC, then it probably shouldn't be an issue.

 

[thestryker]

I think it primarily requires the multiplexer chip be added to the DIMMs. It would be interesting to know if there's a 3rd party multiplexer and if that's what both are using or if either made their own.

I'd be surprised if the multiplexer components were in house:
https://www.renesas.com/en/blogs/re...ts-high-performance-memory-modules-based-ddr5
https://www.renesas.com/en/about/ne...interface-chipset-solutions-second-generation

 

[EzzyB]

It's an industry consortium, so it basically is its member companies.


Maybe, but do we know that, for a fact? It should be completed before there are products on the market. So, the lack of products doesn't imply the standard isn't complete.

I don't know why else they wouldn't have released it. As you say this doesn't seem to be hugely complicated outside the multiplexing.

I think it primarily requires the multiplexer chip be added to the DIMMs. It would be interesting to know if there's a 3rd party multiplexer and if that's what both are using or if either made their own. The point of MCR/MR is that it's supposed to work with commodity DDR5 DRAM, so it wouldn't necessarily be a big lift for Micron to offer this memory, particularly if they can use an existing multiplexer from a 3rd party.

The key point of uncertainty would be the availability of the IP, but if Intel transferred that to JEDEC, then it probably shouldn't be an issue.

Maybe that's it then. We know that both AMD and Intel have provided "specifications" to JEDEC. Maybe a disagreement over which to use or how to resolve differences. Intel not willing to wait until they do?

EDIT: I was thinking further. We do know that Intel got the nod on that new supercomputer project. The one that will use Xeon processors and AMD GPU's. I wonder how much halo projects like this contribute to the decision making.

Sure, the company makes money, but the project, which, at least for a time, will probably end up at the top of the 500 list is probably much more valuable as a marketing tool and Intel can't do it without Xeon 6.

Maybe that's what's, partially at least, behind it.