Accelerating PyTorch Transformers with Intel Sapphire Rapids - part 1

this is pure hardware engineering noise wrapped in some academic PR. intel's claim about sapphire rapids accelerating transformers is technically fine, but the real question is the actual sustained throughput and memory bandwidth when you're running massive models on actual game assets. i don't care about the specs; i care about the latency on our specific deployment pipeline.

if you're just squeezing a few extra percent out of an existing PyTorch setup, it's fine. if you're building a custom accelerator stack, then we need real benchmarks, not just press releases.