1k desktop beats vendor sparse library 474× on Mistral-7B

I've been working on a sparse compute primitive for AI inference. The idea came on a bike ride — I'm a mathematician and serial entrepreneur, not a GPU engineer, and I wanted to see if the matrix math at the heart of every AI model could be restructured to skip unnecessary work while producing identical outputs. Results on my HP All-in-One (i7-1165G7, 4 cores, 64GB, Windows 11):

Mistral-7B real weights, 0% sparsity (fully dense): 127× vs CPU dense, 474× vs CPU sparse, 0.7ms first token vs 49.4ms sparse, 99.2% less energy Llama-2-7B real weights, 0% sparsity: 59× vs dense, 228× vs sparse, 0.8ms first token

On NVIDIA B200 with real HuggingFace weights:

Llama-4 Maverick 400B: 133.5× faster, 99.9% less energy, 52× faster first token DeepSeek-R1 (256 experts): 78.9× faster, 98.7% less energy

The canonical SHA-256 hash appears identically across NVIDIA, AMD, Google TPU, Intel, and Apple Silicon — same math, different silicon, same answer. Independently validated by University of Miami Frost Institute. Open verifier at rolv.ai — runs on any hardware, generates your own baseline hash. No IP in the verifier. Happy to answer technical questions about how it works.