NVIDIA. AMD.
INTEL. NO FAVORITES.
Real benchmark data from a 9-machine homelab fleet. Tested against actual AI workloads — local LLM inference, image generation, video encoding. Published as it runs.
The Hardware
Every benchmark runs on dedicated hardware. No cloud, no shared resources, no synthetic scores.
ROCm · Dual GPU · 64GB Total VRAM
CUDA · Ampere + Blackwell
CUDA · Blackwell Flagship
Cross-Vendor · AMD + NVIDIA
SYCL · oneAPI · 28GB Available
CUDA · Ada Lovelace
First Releases
Starting with Intel Arc Pro B70 — the 32GB VRAM card nobody else is benchmarking seriously.
Intel Arc
Intel Arc
Head-to-Head
How Results Look
Machine-readable benchmark tables in every article. Vendor color-coding makes cross-platform comparisons instant.
| GPU | Model | Backend | Prompt Eval (t/s) | Gen (t/s) | Relative |
|---|---|---|---|---|---|
| Dual R9700 32GB×2 | Qwen3.5-35B-A4B Q8 | ROCm 7.2 | 2,500+ | 65 |
★ |
| Arc Pro B70 28GB | Qwen 35B MoE Q4 | Vulkan | 869 | 39 |
|
| Arc Pro B70 28GB | Gemma 4 31B Q4 | SYCL | 638 | 17 |
|
| RTX 3090 24GB | Mistral Small 24B Q6_K | CUDA | 480 | 36 |
|
| RTX 5080 16GB | Flux1-schnell-fp8 | CUDA | — | 13s/img |
|
Automated Pipeline
WOL to published draft without a human in the loop. Generated from real run data, reviewed before going live.
NVIDIA · AMD · INTEL — No Affiliates, No Favorites