I'm using AMD AI Max 395
The ROCm driver hasn't been released so I couldn't use it properly
While talking with Gemini, when I mentioned this way of doing things is inconvenient, they recommended Lemonade to me
They also recommended something called Vulkan, but Lemonade looked more convenient so I installed it and ran benchmarks
23% improvement
With things like qwen3-30b, since it's an MOE model, it produced 90 tokens, but granite4.1:30b is a dense model so it's just purely full tokens
With 12 tokens, if it accumulates it would be 20% which is significant
It's a RAM bandwidth issue, so for now I think I'll have to be satisfied with this level
