llama.cpp runs faster on E-cores than P-cores in a GPU-offloaded MoE test

dir3ctly · reddit · 2026-07-25

The author benchmarks llama.cpp on a hybrid Intel CPU setup with an RTX 5090, comparing different CPU pinning strategies while offloading a large MoE model.
On their machine, using only E-cores outperforms mixes that include P-cores:
0–23 cores: 19.8 tok/s
E-cores only (12–23): 22.7 tok/s
P-cores plus some E-cores (0–11): 15.6 tok/s
The poster asks why performance gets worse when P-cores are involved, noting that memory bandwidth seems to be the main bottleneck and that they are testing with Docker CPU pinning and CUDA 13.
This is a practical tuning/debugging thread about local inference throughput and heterogeneous CPU behavior.

More from Infra