A user benchmarked the AMD EPYC 9374f against the 9135 for local LLM inference using ik_llama.cpp and Unsloth, finding that the newer chip offers significant decoding speed improvements when configured correctly.
- The 9374f achieved higher token generation speeds (e.g., 12.19 t/s vs 8.68 t/s at N_KV=0) compared to the 9135 when using 48 threads.
- Using 32 or 64 threads resulted in worse performance than the 9135, indicating that thread count is critical for this workload.
- The 9374f showed slightly slower prefill speeds (e.g., 256.90 t/s vs 257.33 t/s) but superior throughput during generation.
The author notes that while the 9374f is better for decoding, it is significantly worse for gaming, suggesting the upgrade may not be worth it for mixed-use systems.