3 ms·
AMD's "HCC" is C++ with templates added onto it. Unlike CUDA, HCC implements all relevant features in C++ Templates alone. If you simply do 64-thread per "warp
by dragontamer 9y ago
AMD's "HCC" is C++ with templates added onto it. Unlike CUDA, HCC implements all relevant features in C++ Templates alone.
If you simply do 64-thread per "warp" (AMD's groups are per-64) and 32KB LDS aka Shared memory per block, you would be able to write portable high-performance code between AMD GPUs and NVidia GPUs.
AMD seems like its a bit behind with regards to GPGPU adoption. But AMD's hardware seems to be a good bit cheaper. You can get HBM2 models at ~$2000 from AMD (Firepro WX9100, which is Vega architecture)
Although... as they say... hardware is cheap. I'm sure most datacenters will prefer the $8000 NVidia V100 instead, because there are more people using that hardware. In particular, its easier to get started with a V100 due to AWS and other cloud-compute offerings.