That is not only wrong, but a little ridiculous.

Yes exactly… those Broadcom Ethernet switches are very fast at 51Tbps. That’s why I don’t understand why we need yet another standard with UALink instead of just modifying the existing Ethernet standard to accommodate this need.

Thanks. It’s my wildly generalized take on what I’ve seen happen in the industry over the last 20 years or so. And based on a decidedly layman’s viewpoint. I would also like to offer a clarification on my point of UAlink being “Intel’s UAlink”. Of course Infinity Fabric is AMD’s interconnect technology and a damn fine one at that. This is why it was chosen for the underpinnings of UAlink. But let’s be honest. Without Intel’s name on the banner page and throwing their support towards the standard UAlink would not have the same traction that it will have otherwise. Yes, I know Cisco is there and HP and Microsoft. But AMD has had those endorsements before for other tech frameworks only to see initial support fade away once Intel pulled their head out of their nether regions and came up with something else. With Intel being in the ranks of support for UAlink along with it being the leader of CXL and UXL and as original member of Ultra Speed Ethernet I think we’re finally seeing a unified approach to compete with Nvidia. I am glad actually to see a great technology developed by AMD as part of this mix.

No need to apologize for lengthy statements. We welcome them.

It’s about load/store, and one implementation of it will run atop Ethernet. Lots more bandwidth in those Broadcom switch ASICs than in the PCI-Express ASICs — at least for now. Perhaps they will accelerate the PCI-Express roadmap–Jas has wanted to do that and skip PCI-Express 6.0 and go right to PCI-Express 7.0, but none of the server and GPU makers were ready.

Basically, AMD and Broadcom are those with THE tech. The rest are along for the ride, at this moment. This is more like when Microsoft, basically, informed Intel that they are going with AMD64 and they better behave or be left behind …

CXL is the heterogeneous memory protocol for connecting CPUs to CPUs, CPUs to GPUs and CPUs to Accelerators that are not GPUs like FPGAs and DSPs. It now contains the work of other heterogeneous protocols such as Gen-Z for heterogeneous communication from rack to rack and CCIX formerly of ARM and OpenCAPI from IBM. Intel spearheaded CXL before all the other consortiums gave over their protocols to the CXL group. You can look at CXL as the logical extension and evolution of PCIe.

You can also look at UALink as the hardware or physical layer to the UXL Foundations software framework. UXL is the foundation set up to promote a unified accelerator framework leveraging Intel’s OneAPI as the alternative to Nvidia”s CUDA. Intel also heads this one up along with Google, ARM, Samsung, Imagination, Qualcomm, Fujitsu and VMware which by the way is owned now by Broadcom who is also part of UAlink. UXL can also be looked at as the successor to AMDs HSA Foundation which was killed off by Lisa Su after she arrived at AMD and killed off the Fusion project. UXL, with the exception of AMD and Google, has pretty much all the other companies that were at one time involved with HSA and for the same reasons that being to have a unified software framework for heterogeneous accelerators whereas in addition UAlink is now the physical side of the former HSA.

So taking all this into account it becomes clear to me at least that the Compute Industrial Complex that’s not Nvidia has chosen Intel heterogeneous hardware links ( CXL ) for everything but GPU to GPU communication and has chosen Intel’s OneAPI as its alternative to CUDA as shown with the UXL Foundation that being the successor to AMDs HSA. And all AMD gets for its pioneering work with HyperTransport and later Fusion and HSA is a pat on the head and the inclusion of its Infinity Fabric for nothing more than heterogeneous GPU to GPU communications.

I mean, that is important and necessary in this age of AI and Global Compute. And also an acknowledgment that Infinity Fabric is superior to other protocols for this purpose. But AMD also needs to acknowledge that once again the industry has rejected its work concerning 3DNow and HSA and ROCm. Lisa Su needs to take a VERY hard look at the Compute Industrial Complex and realize that the only way to compete against Nvidia is to have superior hardware and more importantly superior AND UNIFIED software. AMD makes superior hardware to Intel….full stop. AMD makes shitty software….full stop. AMD needs to rally around Intel’s CXL and OneAPI, leverage the industry support of Infinity Fabric for GPU communication and CEASE development of ROCm and pour those freed up resources and money into further making hardware that runs Intel’s physical and software frameworks better and cheaper than Intel itself. THAT’S….how you compete against Nvidia. Nvidia is the Apple Computer of the Compute Industrial Complex. They make both superior hardware and software in house. The x86 world is the Android version of the Compute Industrial Complex. But by rallying around Intel’s CXL and Intel’s UAlink ( leveraging AMD’s Infinity Fabric ) and Intel’s OneAPI through the UXL Foundation ( the successor to AMDs HSA ) the x86 world might have a chance to competently compete against Nvidia by the end of the decade.