ITADN

[QST] Difference between Fp4 GEMM and Fp4 Ultra Gemm?

#3358Openankutalev 创建于 2026-06-30
question? - Needs Triage
A
ankutalevcommented
Hello! Could you please explain the difference between these two CUTLASS examples: https://github.com/NVIDIA/cutlass/blob/main/examples/72_blackwell_narrow_precision_gemm/72b_blackwell_nvfp4_nvfp4_gemm.cu https://github.com/NVIDIA/cutlass/blob/main/examples/89_sm103_fp4_ultra_gemm/89_sm103_fp4_ultra_gemm.cu At first glance, both seem to use NVFP4 and don’t appear to have underlying differences related to TCGen05. So, what makes them different? Additionally, in the 89_sm103_fp4_ultra_gemm.cu example, there’s a K‑shape tile of size 768, which isn’t mentioned in the Blackwell functionality table (). Is this table still accurate and up to date? Thank you very much for your help!
0 条评论