13 ms·
They're slow, but OK for inference. In practice no one uses AMD/Intel IGPs because no knows about the mlc-llm vulkan backend. llama.cpp is en vogue on the desk
by brucethemoose2 3y ago
They're slow, but OK for inference.
In practice no one uses AMD/Intel IGPs because no knows about the mlc-llm vulkan backend. llama.cpp is en vogue on the desktop, which does not support IGPs outside of Apple, and otherwise people use backends targeted at server GPUs.