Originally posted by atomsymbol
Back when Pascal launched, fp16 was barely a thing. OpenGL had it, but it wasn't much used before deep learning took off. The P100 was the first dGPU to support packed fp16 dot product, but I think it also featured in Gen 9 Intel iGPUs. What the consumer line of Pascal cards had was 4x int8 dot product + accumulate (dp4a), which is great for inferencing.
Comment