xgboost-metal hits 2.56x speedup over 18-core CPU on M5 Pro Mac mini
bingxu_ · x · 2026-11-06
Pratik Gandhi benchmarked Bojan Tunguz's xgboost-metal (XGBoost accelerated via Apple Silicon GPU/Metal) on an M5 Pro Mac mini:
- 5M rows × 64 features: 13.7s on Metal GPU vs 35.0s on all 18 CPU cores — 2.56x faster
- 1M rows: 3.4s vs 7.8s, a 2.29x speedup
- Accuracy and AUC matched the CPU version to 4+ decimals
- Config: hist tree method, 500 trees, maxdepth 8, learning rate 0.1
Tunguz notes even smaller GPUs like the Mac mini see significant speedups, with more gains planned in upcoming releases. Reposter bingxu went further, claiming AMD and Intel CPUs are "dinosaurs" and that his SwarmOS project (his team used Llama 3 two years ago to generate valid GPU kernels) will reshape hardware design even if frontier models get restricted.
More from Infra
- Google to invest $15 billion in Malaysia for data center campuses in 2027-2028 — GabGarrett · 2026-11-06
- Ripple-65B on one RTX PRO 6000: 110-120 tok/s on code in local agent tests — DZhang50 · 2026-11-06
- NVIDIA open-sources cuRoboV2-Fast: robot motion planning cut from 100ms to 4.4ms — ankurhandos · 2026-11-06
- DeepSeek-V4.1-Flash runs on one RTX 3090 at 41 tok/s with Delta Inference Engine v0.93 — musicurgy · 2026-11-06
- India's TRAI recommends direct-to-device satellite service, a win for Starlink Direct to Cell — XFreeze · 2026-11-06
- Developer still can't rent A10g instances, suggesting GPU demand remains tight — generativist · 2026-11-06