My personal predictions:

#1: Don't expect to see it until after AMD or someone else makes a product that actually manages to compete against Nvidia. Nvidia is pretty hostile to not running their low-level software stack, and Oxide is all about the legacy of Solaris.

#2: AMD MI400 or relative will be an extra chipset on future server motherboards (not a separate PCI card). Simultaneously, the boundary between "CPU vector processing" and "GPU used for transformers" will blur, and the chipsets will slowly merge into chiplets in one package.

#3: AMD MI400 and such AI accelerators will be primarily sold as full racks with its own custom "networking" (UALink switch), and the actual host CPU on those devices will be lower specs and mostly relegated to setup and metrics of the AI work, much like storage and networking appliances are built, AI workload will not even pass through the host CPU. I'm not sure Oxide can compete in that world. The "business logic CPUs" will reside in a different rack.

▲

zozbot234 3 days ago | parent [-]

> AMD MI400 or relative will be an extra chipset on future server motherboards (not a separate PCI card). Simultaneously, the boundary between "CPU vector processing" and "GPU used for transformers" will blur, and the chipsets will slowly merge into chiplets in one package.

Isn't this just an iGPU? They generally have much lower performance than GPU's sitting on a dedicated card.

	▲	yencabulator 3 days ago \| parent [-]
		It's built to be deployed in units of racks, so that "i" is a little funny. It's not a little wart on the side of a general-purpose CPU. 432 GB memory at 19.6 TB/s are the claimed specs. For one package. Now fill a rack with 72 of those. (If only AMD could execute on the software side...)