Meta: Co-Design for AMD GPUs Delivers Up to 1.5x Training Throughput vs Direct NVIDIA Ports
burny_tech · x · 2026-10-20
Meta's new paper reports the first hyperscaler account of production-scale DLRM training on AMD GPUs, matching NVIDIA B200 performance with MI350X. Key points:
- Hardware-model co-design beats direct porting: revising model structure and software recipes to exploit MI350X's 1.6x HBM capacity over B200 — bigger batches, less activation checkpointing, embedding re-sharding.
- Analytical performance model enables near-zero-warmup optimal kernel selection across diverse GPU generations.
- Agentic workflows rewrite microarchitecture-aware kernels and custom shims, slashing the porting barrier from the CUDA ecosystem — yielding up to 1.5x end-to-end training throughput vs the direct port baseline.
Contributions are upstreamed to PyTorch, Triton, and ROCm — a practical blueprint for multi-vendor GPU strategies.
More from Infra
- Q-CTRL: over 99.5% of qubits sit idle in one fault-tolerant Shor's algorithm blueprint — MJBiercuk · 2026-10-20
- Nscale and Anthropic reportedly in talks for another $45 billion compute deal — TechCrunch AI · 2026-10-20
- Qwen3.8 quant testing on 5090/Strix Halo: Gufo engine fixes tool-calling failures — awitod · 2026-10-20
- SGLang Summit sponsors jump from 11 to 36 in a week; Lip-Bu Tan, Lilian Weng to speak — satnam6502 · 2026-10-20
- Fused linear DPO loss merged into NVIDIA TileGym: 7.2x less memory, 1.27x faster — blelbach · 2026-10-20
- If GitHub goes down, how does GitHub push code to GitHub to bring GitHub back up? — VeryWellVersed · 2026-10-20