ptrendx commented on issue #13650: speed drop when set `multi_precision=True` URL: https://github.com/apache/incubator-mxnet/issues/13650#issuecomment-448462206 It's hard to give a definitive answer without looking at the profile, but the most probable explanation is as follows: - `local` kvstore uses CPU to perform the reduction and update of the parameters. CPU can not natively handle FP16 data (there are instructions in Ivy Bridge+ that help with casting to/from fp16, but I'm not sure they would be used there) and so that step is very slow - therefore the update step when not using multiprecision option is probably already pretty close to become a bottleneck in your training. Multiprecision version needs to do more work - at the very least read and save fp32 copy of weights in the case of SGD which has a special kernel for it, but in the case of NAG there is no special kernel, so the overhead is bigger. That is probably why you see the impact - the cost can no longer be hidden behind the computation on the GPU.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
