Comment by Vvector
Chip speed isn't as important as good software
Chip speed isn't as important as good software
The average consumer uses llama.cpp. So here is your list of kernels: https://github.com/ggml-org/llama.cpp/tree/master/ggml/src/g...
And here is pretty damning evidence that you're full of shit: https://github.com/ggml-org/llama.cpp/blob/master/ggml/src/g...
The ggml-hip backend references the ggml-cuda kernels. The "software is the same" (as in, it is CUDA) and yet AMD is still behind.
The software is the same, AMD is not doing its own LLMs.