AMD Instinct MI325X inference latency for batch-1 serving

উত্তর

AMD Instinct MI325X performance headline: 1,307 FP16 TFLOPS, 163.4 FP32 TFLOPS, 6,000 GB/s bandwidth, 256 GB VRAM.

Converted into practical benchmarks: model training a 7B-parameter LLM in FP16 with reasonable batch sizes typically saturates compute before bandwidth; real-time serving on the same model is usually bandwidth-bound and tracks the 6,000 GB/s figure. Diffusion image generation benchmarks sit between the two — compute-heavy steps utilise tensor cores well, while attention blocks still touch bandwidth.

The cheapest AMD Instinct MI325X cloud access right now is on Vultr at $2.00/hr.