Closed dashesy closed 6 years ago
Hi @dashesy , it takes about 1m45s to build it on my box. Yes, it is slower to build than BVLC because of fp16 support - you actually build both versions, fp32 and fp16.
Is there a way to only compile for float32?
Not at this moment, sorry
The compile time of NVIDIA/caffe (Cuda9 and CUDNN 7) is much slower than BVLC/caffe. In both cases I compile with "All" architectures because I do not know which one may be running on the target cluster. Is this expected because of new architectures?