AMD Instinct MI300X ☆

The MI300X stacked 8 CDNA3 compute chiplets on 4 base dies — 153 billion transistors, more than any GPU before it — with 192 GB of HBM3 when the H100 had 80. That memory advantage won real deployments: Azure ran GPT-4-class serving on it, Meta served Llama on it, and it proved the ROCm software stack at scale. The knock: software maturity, not silicon, remained AMD's gap.