Sam
@spatters.bsky.social
Computer stuff and skiing. Occasionally pizza.
Wondering why A100 tensor cores have 2x the flops on FP16 and BF16 that they do on TF32.
Based on mantissa size I would have guessed FP16 and TF32 should be the same, and BF16 faster, so must be missing something.
Based on mantissa size I would have guessed FP16 and TF32 should be the same, and BF16 faster, so must be missing something.
May 3, 2023 at 4:48 PM
Wondering why A100 tensor cores have 2x the flops on FP16 and BF16 that they do on TF32.
Based on mantissa size I would have guessed FP16 and TF32 should be the same, and BF16 faster, so must be missing something.
Based on mantissa size I would have guessed FP16 and TF32 should be the same, and BF16 faster, so must be missing something.