Intel’s Ponte Vecchio: chiplets gone crazy

Intel is a newcomer to the world of discrete graphics cards, and the company’s Xe architecture is driving its effort to establish itself alongside AMD and Nvidia. We’ve seen Xe variants serve in integrated GPUs and midrange discrete cards, but Intel’s not stopping there. Their GPU ambitions extend to the datacenter and supercomputing markets. That’s where Ponte Vecchio (PVC) comes in.

Like other compute-oriented GPUs, PVC goes wide and slow. High memory bandwidth and FP64 throughput differentiate it from client architectures, which emphasize FP32 throughput and use caching to reduce memory bandwidth demands. Compared to Nvidia’s H100 and AMD’s MI210, PVC stands out because it lacks fixed function graphics hardware. H100 and MI210 still have some form of texture units, but PVC doesn’t have any at all. Combine that with its lack of display outputs, and calling PVC a GPU is pretty funny. It’s really a giant, parallel processor that happens to be programmed in the same way you’d program a GPU for compute.

Another great feature from Chips and Cheese. Speaking of Intel – the company also unveiled that Meteor Lake CPUs are coming to the desktop in 2024.