Jaw droppingly kick ass tech. Gobs of bandwidth, incredible fabric acceleration. 22B transistors, 400Gbe/infiniband (maybe x4, hard to tell), PCIe 5.0 x32, 16x

Jaw droppingly kick ass tech. Gobs of bandwidth, incredible fabric acceleration. 22B transistors, 400Gbe/infiniband (maybe x4, hard to tell), PCIe 5.0 x32, 16x Cortex-A78 ARM cores, datapath accelerator with 16 cores 256 hardware threads. Via coverage from August 2021 of BlueField-3: https://www.servethehome.com/nvidia-bluefield-3-dpu-architecture-at-hot-chips-33/ https://www.servethehome.com/nvidia-bluefield-3-dpu-architec... It could use a couple drops of AWS Nitro like acceleration, but so far that kind of hardware virtualization hasn't really shown up on the broader scene. But... RDMA, oh yes! That said, AMD had acquired SeaMicro who had an incredibly sweet Nitro-like virtualized-peripheral system, which powered some of the sickest most awesome ultra-dense systems built. All quite obsolete now, but it felt like where we ought to go. The fabric is so the thing. More bandwidth, smarter bandwidth/connectivity across systems. CXL has killed/merged-in all competitors, and 3.0 finally actually does what the much better specs (Gen-Z) did semi-somewhat adequately (be a semi-long-distance capable fabric, rather than a fancy peripheral bus). I have a lot of hope & trust Intel's going to do great with their silicon photonics. AMD has been shipping chiplets nicely, but Intel has been shipping far more advanced EIMB, Foveros (Lakefield was so under-rated for so many reasons), and photonics for years now. I hope AMD can stay competitive. Epyc is just kicking ass everywhere, but I really don't see AMD as having any plans for connectivity. I've heard some recent rumblings that they want to start shipping CXL even on consumer hardware, and that would be a great place to get a strong lead, a big selling point, and to get competency. I'm still a little sad, after so many years of AMD being the open HyperTransport company that wanted to play with others- it never materialized, and they gave up. Now the follow-up Infinity Fabric (IF) is all closed- it's unclear what exactly IF really is or does or whether it's just marketing speak... does the IF on CPUs really resemble & work similar to the IF on the new RDNA3 gpus? IF would be interesting as a more fabric-like system, but it's an internal implementation detail we know nothing about, and that's unfortunate for AMD. I'm hoping for a very interesting 2025+ as newer CXL-based RDMA like systems & peripherals & virtualizations emerge. Having very fancy flash drives and network resources that are fabric-attached, having systems doing next-gen RDMA to each other is very compelling. I hope hardware like this can have a role, but I also hope a lot of Intel e-core like acceleration gets thrown into regular core complexes as well & competes, integrated versus accelerators, each trying.