Comment by karmakaze

I really appreciate seeing these dense model numbers. For a large unified memory system though I expect that MoE numbers are what people are more interested in.

These numbers could and should get much better. As an example I can run Qwen3.8-27B-MXFP4 (W4A8) on 2x AMD R9700 that gets 260+ tokens/sec to start and slows down to ~110 tokens/sec over 128k context and can do the max 256k. These are for batch size 1 and throughput goes higher with batching. This is due to speculative decoding, efficient all-reduce inter-gpu compression, and custom GEMM kernels for the specific hardware. Note each R9700 only has 644 GB/s memory bandwidth.