Gpu memory fraction impact tuning result?

There is per_process_gpu_memory_fraction in tensorflow to limit the max gpu memory used by each process, but does not find such configuration if using TVM runtime
Because the module compiled by TVM used the constant gpu memroy, so no need to add the gpu memory fraction?
And does the gpu memory size impact the tuning result?More gpu memory lead to better tuning perf of the same shape conv2d?

1 Like