News for hackers. Stuff that matters.

A comment on Caffe2 adds 16 bit floating point training support on the NVIDIA Volta platform.

Re:Caffe2 adds 16 bit floating point training support on the... by mappu on Thursday May 11, 2017 @11:01PM (#14321113)
Are numbers available for FP16 on P100 or FP32 on V100? It would make for a more direct comparison.

EDIT: Nvidia's advertised TFLOPS are:

         FP16  FP32  FP64
    V100 30    15    8.5
    P100 21.2  10.6  5.3
    K40  4.29  4.29  1.43
[ Parent | on HN ]
Re:Caffe2 adds 16 bit floating point training support on the... by Tom1971 on Friday May 12, 2017 @01:51AM (#14321712)
Your table doesn't include the new "tensor core" TFLOPS of V100.

That's a core that does 4x4 FP16 matrix multiplication + 4x4 FP32 accumulation in one go.

That's where V100 gets its boost, up to 120 TFLOPS.

[ Parent | on HN ]
Re:Caffe2 adds 16 bit floating point training support on the... by bsprings on Friday May 12, 2017 @02:08AM (#14321771)
Tensor Cores: 120 TFLOP/s mixed-precision (peak). Typo in your table: V100 FP64 is 7.5 TFLOP/s.
[ Parent | on HN ]