第 1 則更新，共 1 則最新第 1 天
AMD's Helios Lands 72 GPUs in One Rack, and Four Hyperscalers Say Yes
Microsoft joins Meta, OpenAI, and Oracle as AMD ships an open-standards challenger to Nvidia's grip on frontier AI.
最初掌握的資訊
- AMD formally launched Helios on July 20, 2026.
- Microsoft will deploy Helios racks inside Azure.
- Meta, OpenAI, and Oracle are also named early customers.
- Each rack holds 72 MI455X GPUs and 31 TB of HBM4, with 2.9 exaflops FP4 inference and 1.4 exaflops FP8 training.
當時尚無定論
- Per-rack price, power draw, CPU count per rack, headcount on the program, and AMD's own revenue projection.
- Independent benchmarks comparing Helios to Nvidia's Vera Rubin on real frontier model training.
我們說過將關注的
- Whether Microsoft, Meta, OpenAI, and Oracle move from announcement to volume orders in 2027.
- How ROCm holds up against CUDA on the workloads hyperscalers actually run.
