Intel, AMD, and ARM have joined forces to develop a universal interconnect standard for Chiplets

https://www.anandtech.com/show/1728...d-setting-standards-for-the-chiplet-ecosystem

If you ever needed a sign of the end times, it would be Intel, AMD, ARM, Samsung, and TSMC joining forces to develop a standard.
Probably done to help kerb the chip manufacturing crisis. If everyone follows a standard, then part of the process can be reused between chips, and competing chips can be used in the same application with fewer changes.

This means they can make and sell more chips, although it doesn't necessarily translate into lower prices. Probably will be several years before we see anything from this, if we ever do.
 
Probably done to help kerb the chip manufacturing crisis. If everyone follows a standard, then part of the process can be reused between chips, and competing chips can be used in the same application with fewer changes.

This means they can make and sell more chips, although it doesn't necessarily translate into lower prices. Probably will be several years before we see anything from this, if we ever do.
They've been collaborating on it for the last 5 years apparently, it's now reached a stage where they are announcing it as they have completed the 1.0 specifications.
More details can be found here:
https://www.uciexpress.org/membership

But yes they state their primary goal is to make manufacturing easier and simplify design and development as most chiplet based packages are already using a mix and match of parts from various vendors on different processes.
 
Good. There's really no proprietary value here, buzzwords notwithstanding.
 
Is it only me that finds it peculiar that nvidia is not in that group (so far)?
 
Is it only me that finds it peculiar that nvidia is not in that group (so far)?
Not really what would they have to offer at this stage? Samsung, ARM, and TSMC are members so they benefit by proxy. And had the ARM deal gone through they would have been, and while the ARM deal was being negotiated would it have gone through that would have granted them a larger representation on the board 2 seats to everybody else’s one.
 
Nope. I'm certain they are harder to work with.
Their really not that hard to actually deal with, what they are however is unneeded. They currently have no significant experience working with chiplets to a degree where they could meaningfully offer up any unique insights. They will simultaneously get access to any of the work they do via their work with TSMC, Samsung, and ARM. There are also no benefits for this being proprietary by any means and any of the proprietary tech they have developed can then just remain quietly in house. This is simply a case where they have nothing to offer up from either a design or manufacturing perspective and will probably not look to join until they have a design or two under their belt, probably in time for them to begin ratifying version 2.0 of the specification.
 
Not really what would they have to offer at this stage? Samsung, ARM, and TSMC are members so they benefit by proxy. And had the ARM deal gone through they would have been, and while the ARM deal was being negotiated would it have gone through that would have granted them a larger representation on the board 2 seats to everybody else’s one.
"The promoter members of UCIe are AMD, Arm, Advanced Semiconductor Engineering, Inc. (ASE), Google Cloud, Intel, Meta [AKA Facebook], Microsoft, Qualcomm, Samsung, and Taiwan Semiconductor Manufacturing Company. [TSMC]"
https://www.uciexpress.org/membership

"In short, it’s a who’s who of many of the big players in the chip industry (be them producers or consumers), but the promoters are looking for more members. Pragmatically, the more buy-in the standard gets the more effective and widely adopted it will be, but the group also benefits from the opinions of additional firms, and understanding what their compute needs are."
https://www.anandtech.com/show/1728...d-setting-standards-for-the-chiplet-ecosystem (original link)

From that standpoint I can see why one might be surprised to see NVIDIA absent from the list.

What I find interesting is that the UCIe membership page is titled "My Site". ;)
 
"The promoter members of UCIe are AMD, Arm, Advanced Semiconductor Engineering, Inc. (ASE), Google Cloud, Intel, Meta [AKA Facebook], Microsoft, Qualcomm, Samsung, and Taiwan Semiconductor Manufacturing Company. [TSMC]"
https://www.uciexpress.org/membership

"In short, it’s a who’s who of many of the big players in the chip industry (be them producers or consumers), but the promoters are looking for more members. Pragmatically, the more buy-in the standard gets the more effective and widely adopted it will be, but the group also benefits from the opinions of additional firms, and understanding what their compute needs are."
https://www.anandtech.com/show/1728...d-setting-standards-for-the-chiplet-ecosystem (original link)

From that standpoint I can see why one might be surprised to see NVIDIA absent from the list.

What I find interesting is that the UCIe membership page is titled "My Site". ;)
I suppose, but NVidia may not yet see a benefit to it, or they may have been working on it through ARM as NVidia and ARM have been collaborating on many things for a few years now even before the sales talks started. There are a number of reasons for them to not be on that list yet, though I would be surprised if they didn't end up on it sooner than not.
 
I personally think that they shouldn't do this. I think when it comes to chiplet or MCM designs AMD has an advantage. Sharing that tech with their competitors would cancel out that advantage. Now, I don't think AMD would give them everything and just enough to still hold an advantage but AMD needs to thread lightly with this. What do I know though. Just an opinion on something that smarter people probably have figured out already.
 
Last edited:
I personally think that they shouldn't do this. I think when it comes to chiplet or MCM designs AMD has an advantage. Sharing that tech with their competitors would cancel out that disadvantage. Now, I don't think AMD would give them everything and just enough to still hold an advantage but AMD needs to thread lightly with this. What do I know though. Just an opinion on something that smarter people probably have figured out already.
Not really, but sort of AMD has the better connection designs for CPU’s and maybe GPU’s that’s yet to be seen. Intel has been using it for longer just for different things, memory controllers, networking components, and much of their custom enterprise solutions where they not only work with different chips but different processes as well.

In AMD’s case how much of their MCM interconnects they designed and how much of it TSMC developed is a crapshoot much of it was jointly done. TSMC uses much of that same tech for many of the custom ARM cores they print for other vendors who had the same issues.

There’s always going to be room for a proprietary implementation for special silicon, but there is a very large and growing need for an open spec as well as custom hardware becomes the norm. SoC’s do too good a job and we are going to see a lot more of them leaving the professional work space for the consumer ones and to do that you need at least a rudimentary standard they can all follow.
 
Their really not that hard to actually deal with, what they are however is unneeded. They currently have no significant experience working with chiplets to a degree where they could meaningfully offer up any unique insights. They will simultaneously get access to any of the work they do via their work with TSMC, Samsung, and ARM. There are also no benefits for this being proprietary by any means and any of the proprietary tech they have developed can then just remain quietly in house. This is simply a case where they have nothing to offer up from either a design or manufacturing perspective and will probably not look to join until they have a design or two under their belt, probably in time for them to begin ratifying version 2.0 of the specification.

I didn't mean chiplets. I meant nVidia.
 
They've been collaborating on it for the last 5 years apparently, it's now reached a stage where they are announcing it as they have completed the 1.0 specifications.
More details can be found here:
https://www.uciexpress.org/membership

But yes they state their primary goal is to make manufacturing easier and simplify design and development as most chiplet based packages are already using a mix and match of parts from various vendors on different processes.
Mix and match of parts from various vendors... That sounds like a really good thing to me. Otherwise, you have a potential for a monopoly.

I suppose this works if the spec is open like USB and any vendor has the potential to design?
 
Mix and match of parts from various vendors... That sounds like a really good thing to me. Otherwise, you have a potential for a monopoly.

I suppose this works if the spec is open like USB and any vendor has the potential to design?
Well with this you could in theory use an Intel memory controller done on 14nm, paired with some Intel efficiency cores done on Intel 7, and some ARM cores done on TSMC 5 then patch that all in with an AMD GPU done on TSMC 7 and package that all under one lid. Now you potentially have a super low power device that when it needs access to some x86 instructions it doesn’t have to emulate them. For most legacy office applications a pair of Intel efficiency cores is more than enough to tackle that.

If I understand it correctly that is, I may be way out to lunch on the actual workings of the spec.
 
Not really, but sort of AMD has the better connection designs for CPU’s and maybe GPU’s that’s yet to be seen. Intel has been using it for longer just for different things, memory controllers, networking components, and much of their custom enterprise solutions where they not only work with different chips but different processes as well.

This is probably a sign that AMD doesn't think their interconnect tech is going to stay at the front much longer. Another sign is them buying Xilinx. It kind of has to be a shared standard going forward unless they're planning to only do custom SOCs. Beyond Microsoft and Sony and a few others, who can count on having that kind of money?

I doubt Intel or AMD is going to tap each other for chiplets, but I could see them both selling chiplets to other companies building their own Lego SOCs.
 
This is probably a sign that AMD doesn't think their interconnect tech is going to stay at the front much longer. Another sign is them buying Xilinx. It kind of has to be a shared standard going forward unless they're planning to only do custom SOCs. Beyond Microsoft and Sony and a few others, who can count on having that kind of money?

I doubt Intel or AMD is going to tap each other for chiplets, but I could see them both selling chiplets to other companies building their own Lego SOCs.
AMD hasn't been shy about the fact its interconnects and caching structures were jointly developed with TSMC, that collaboration may have had some degree of exclusivity but with TSMC working alongside just about everybody, that sort of stuff is going to get shared around. Intel needs to play ball here because they are looking to get into spinning up fabs for 3'rd parties to use and they have been taking notes from TSMC and Samsung on how to operate Fabs in that manner, here they don't have the luxury of being the odd man out and not supporting these standards because it would only hurt their business practices here and limit what sorts of contracts they could take while increasing costs, so getting in on the ground floor and making sure the spec is compatible with their processes is important if they are going to be a viable fab for others.

Lego SOCs, I like that term... Are going to become more of a thing, it has become extremely apparent to all the major players that SOC solutions are incredibly efficient, cheaper to produce, and offer significant performance benefits. It's also become painfully apparent to them that they can not stand up against apple when it comes to theirs, creating some means for each vendor to more easily interconnect is their only chance to match up with out spending very large amounts of money branching into new fields and competing on custom platforms which is a loosing game for each of them and they know it, they can't really afford to fragment their markets any more than they already are because that hurts adoption which ultimately leads to failed launches.
 
Minor necro post here.

Some of you may recall that Nvidia's conspicuous absence from the UCIe Consortium's board members had led to some questions and speculation back when it was first announced.
We can all finally sleep at night, as the matter has been settled: UCIe Consortium Incorporates, Adds NVIDIA and Alibaba As Members

"The UCIe Consortium Board includes founding members Advanced Semiconductor Engineering, Inc. (ASE), AMD, Arm, Google Cloud, Intel Corporation, Meta, Microsoft Corporation, Qualcomm Incorporated, Samsung Electronics, and Taiwan Semiconductor Manufacturing Company, and newly elected members, Alibaba and NVIDIA."
https://www.uciexpress.org/_files/ugd/0c1418_93535a298a1444b3a39be31f878c68f6.pdf
 
Back
Top