How to check tensor core has been used


(Weiqiangguo) #1

OS: Win10, Cuda10 VS2015 RTX2080 TI
I use pytorch C++.

bool DeepReconImpl::Initial(char* strModPath,int iDeviceID)
{
at::Device m_Device = at::Device(at::kCUDA, iDeviceID);

std::filebuf inModelFile;
inModelFile.open(strModPath, std::ios::in | std::ios::binary);
std::istream inModel(&inModelFile);

m_pNet = torch::jit::load(inModel);
m_pNet->to(m_Device, false);

inModelFile.close();
return true;

}

When test cases are running, the GPU Cpmpute_0 is used at 90% from windows task manger.
But I don’t know if RXT 2080ti tensor cores are been used.


#2

Tensor cores are used if the inputs are float16, the shapes are multiples of 8 and it’s a matmul call.


(Weiqiangguo) #3

Is there a tool for judging tensor core has been used? GPU-Z? or some tools?


#4

no there isn’t such a tool.