Caffe2 is itself designed for performance, and is significantly faster than non-graph-based frameworks such as PyTorch, Chainer, DyNet...
New TensorFlow benchmarks: faster training than Caffe2; near-linear speedup with # of GPUs https://www.tensorflow.org/performance/benchmarks … https://blogs.nvidia.com/blog/2017/04/18/caffe2/ …pic.twitter.com/1bG4uAHEuZ
-
-
-
Have you tried this on phi? If it scales linearly you could put 128 cores to work
End of conversation
New conversation -
-
-
Great! How it compares to CNTK which boasts for its scalability?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Looks even better than linear! Any multi-machines benchmarks?
-
Maybe not better but pretty close :)
End of conversation
New conversation -
-
-
you make a strong claim (TF faster than C2). The benchmarks are changed because of a small bug, and now TF not faster than C2.
-
Why not tweet about this change too? Kudos to TF team for not pitting TF vs C2, but you did, so it seems responsible to fix misinformation
- Show replies
New conversation -
-
-
Not in distributed setups - these are (some of) the good points for caffe2, tf (and even PyTorch has 8 GPU). Not so easy in Theano right now
-
Of course, controlling and planning the hardware is a key part of distributed training but still - pushing multi GPU speed is important IMO
End of conversation
New conversation -
-
-
Also, while performance is important, the ecosystem, support tooling, and deployment are often more important—glad for competition there.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.