Benchmarks
Benchmarks from my RTX PRO 5000 and Strix Halo setups.
How to read a benchmark page here
- The tested environment block states the hardware, OS, drivers, and exact software versions. If your stack differs, the numbers are a reference point, not a prediction.
- Interpretations are scoped to the models, quantizations, and workloads actually run — and say so. A ranking on one model is not a claim about every model.
- The tables are the complete record — mean and spread per test, plus the settings behind them. Raw harness output stays on the test machine and is shared out-of-band on request; it is not hosted.
- Pages are pinned to the versions tested, not to "current" software — upstream moves faster than test runs do.
All benchmarks
- RTX PRO 5000 72 GB + 48 GB: Local LLM Benchmark
llama.cpp benchmarks on my 72 GB + 48 GB RTX PRO 5000 setup, comparing single-card runs, layer splits, and tensor splits.
- ROCm vs Vulkan on AMD Strix Halo: llama.cpp Benchmarks
ROCm and Vulkan llama.cpp results on Strix Halo, comparing a dense model and an MoE model, with additional speculative-decoding tests.
- GPU-to-GPU P2P Bandwidth on Z890 (Dual RTX PRO 5000)
Measured GPU-to-GPU copy bandwidth on my Z890 system, comparing direct peer copies with transfers staged through host memory.