FP8

An 8-bit floating-point format that roughly doubles throughput and halves memory versus 16-bit for AI workloads.

Current numbers

~2.9 EF FP4AMD Helios (72× MI455X) rack-scale FP4 (1.4 EF FP8), 31 TB HBM4, ORW double-wideas of 2026-10 · register ↗

~29% lower (B200 1.41x)MI355X (ROCm, SGLang, FP8) cost per million tokens vs B200, GLM-5 on one node, best point, May 2026as of 2026-05 · register ↗

192 GBHBM3E per Ironwood (TPU v7) chip @ 7.37 TB/s; 4,614 FP8 TFLOPS; 6x memory vs Trilliumas of 2026-03 (GA) · register ↗