Comparing GPU programmable platforms on NVIDIA and AMD

So far, we have explored the scope of computing on NVIDIA and AMD GPUs through two separate chapters. Now, let's specifically look into the comparisons between their respective APIs:

NVIDIA CUDA	AMD ROCm
The API is called Compute Unified Device Architecture	The API is called Radeon Open Compute platform
Proprietary	Open source
Released in 2007	Released in 2016
Wider support	Still under adoption and very actively catching up
Significant number of programmable libraries	Fewer libraries than CUDA but active ongoing development
Cannot be used with non-NVIDIA devices	Cross-platform independence due to open standards
CUDA-C language being used	HIP for cross-platform; HC for AMD GPUs
.cu extension used for files	.cpp extension used for files
Non-portable	CUDA code portability possible with HIP
OpenCL compatible	Also OpenCL compatible
Significant progress in machine learning	Steady progress in machine learning

 

It is interesting to note that C++ programmers learning GPU computing for the first time will still be using the .cpp extension with AMD ROCm. So, in terms of familiarity, ROCm looks like a better option. NVIDIA GPU owners can opt for HIPCC, whereas AMD GPU owners can go for HCC.