#torchao

Large Language Models

Intel Core Ultra 3 Delivers 58ms LLM Tokens via PyTorch 2.10 — But Is It Enough?

58 milliseconds to spit out the first token from a Qwen model. Intel's Core Ultra Series 3, juiced by PyTorch 2.10 and TorchAO, claims it's ready for prime-time AI on your laptop — but let's poke holes in the hype.

MXFP8 MoE Training: 1.3x Speedup, But Skepticism Lingers

MXFP8 just turbocharged MoE training by 30% on massive GB200 setups. Equivalent quality to BF16? Sure. But let's poke holes in the hype.