A drop-in matmul operator that delivers up to 106× faster AI inference and up to 99% less energy on the same hardware — with bit-identical output.
Go to rolv.ai, sign in, pick a model from the dropdown, click run. The compute happens on our server via the public benchmark API (see rolvai/benchmark on HuggingFace). A SHA-256-signed result lands in your inbox with per-case speedup, energy reduction, correctness check, and a run hash bound to your hardware fingerprint.
Works on a laptop. Works on a Chromebook. Works on a phone. The receipt is cryptographically tied to your run — it cannot be copied from another benchmark, and cannot be fabricated without actually running.
One-time result email. No newsletter. No follow-up sequence.
NVIDIA B200, BF16, TF32 on, 1,000 iterations:
| Model | Natural sparsity | vs cuBLAS | vs cuSPARSE | Energy reduction |
|---|---|---|---|---|
| Llama-4-Scout | 93.8% | 4.75× | 103× | 79% |
| Mixtral-8×7B | 75.0% | 1.86× | 109× | 46% |
| Qwen3-30B-A3B | 93.8% | 3.43× | 32× | 71% |
| OLMoE-1B-7B (H200) | 87.5% | 2.49× | 43× | 60% |
Intel i7 laptop (4 cores, 68 GB RAM, MKL baseline):
| Model / Layer | Sparsity | vs MKL |
|---|---|---|
Llama-3.2-1B down_proj |
99% | 106.65× |
Qwen2.5-7B gate_proj |
95% | 59.70× |
Mistral-7B q_proj |
95% | 21.45× |
Full per-case data with SHA-256 hashes: rolv.ai/rolv_benchmarks.pdf
Modern AI weight matrices are mostly zero. In a Mixture-of-Experts model like Mixtral or DeepSeek-V3, 75–97% of weights are architecturally inactive for any given token — guaranteed by the router, known before computation starts. Standard libraries compute all of them anyway.
ROLV identifies the non-zero structure at load time and restricts computation to live elements only. The operation uses the same underlying BLAS / tensor-core primitive on a matrix proportional to the non-zero fraction — but on a single contiguous submatrix rather than indexed scatter-gather. Results are placed into the correct positions of the full output tensor. Final output is bit-identical to the full dense operation.
Inner mechanism is Patent Pending.
Every benchmark case is produced with five independent verification layers:
- Real HuggingFace weights — downloaded from public repositories, no synthetic matrices
- Vendor baseline — Intel MKL on CPU, cuBLAS on GPU, cuSPARSE at high sparsity
- Four SHA-256 hashes per case — input matrix, input vector, vendor output, ROLV output
- Perturbation test — one weight altered by 10⁻³, output hash must change (rules out cached answers)
- Signed run hash — SHA-256 over speedup, timestamp, and hardware fingerprint
ATOL = 0.05 on column-normalised fp64. The correctness check and the speed measurement are the same execution — work cannot be skipped to game the clock without also failing correctness.
1,684 / 1,684 GPU PASS · 332 / 332 CPU PASS
Independently validated by the University of Miami Frost Institute for Data Science and Computing — bit-identical SHA-256 hashes across CPU, GPU, and TPU. No commercial relationship. Validation letter.
Validated today: NVIDIA (B200, H200, H100, A100, RTX series, T4, V100) · AMD (MI300X, MI250X, RX 7900) · Intel CPU (MKL, AVX-512) · AMD EPYC · ARM Neoverse · Apple Silicon (M1–M4 Pro) · Google TPU (v4, v5)
Framework support: PyTorch · JAX · TensorFlow · ONNX Runtime · TensorRT · vLLM · HuggingFace Transformers
Works with: BF16, FP16, FP32, INT8 checkpoints. No retraining. No re-quantisation.
The public benchmark at rolv.ai proves the output. To run ROLV on your own hardware, against your own models, in your own environment, two NDA-gated tiers are available:
Secure Container — hardware-locked Docker container. RolvKey™-authenticated. Processor fingerprint binding at first run; will not execute on any other machine. Optional Intel SGX hardware encryption for regulated environments. Evaluation licence + NDA required.
Direct Hardware — single authenticated file for bare-metal servers and air-gapped environments where Docker is not permitted. Processor-bound binary with live heartbeat attestation. Evaluation licence + NDA required.
Both tiers return cryptographically signed per-run results bound to your processor fingerprint.
Contact: rolv@rolv.ai
If you use ROLV in research, please cite:
@software{heggenhougen_rolv_2026,
author = {Heggenhougen, Rolv E.},
title = {ROLV Primitive©: A Universal Compute Primitive for Sparse AI Inference},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.19221455},
url = {https://rolv.ai}
}GitHub's "Cite this repository" button (top-right of this page) pulls the same information from CITATION.cff.
- Live benchmark: rolv.ai
- Paper: Zenodo 10.5281/zenodo.19221455
- Independent validation: rolv.ai/validation
- HuggingFace: huggingface.co/rolvai
- Benchmark API Space: huggingface.co/spaces/rolvai/benchmark
- Substack: rolv.substack.com
Contact: rolv@rolv.ai — a real person reads every message.
This repository (documentation, citation, and project metadata) is licensed under Apache 2.0. See LICENSE.
The ROLV Primitive© implementation is Patent Pending and licensed separately under NDA for on-premise evaluation. The Apache licence on this repository applies only to the documentation and project files here — it does not grant any licence to the underlying algorithm or implementation.
ROLV LLC · Fort Lauderdale, FL · Patent Pending · ROLV Primitive© · RSMT™ · ROLVswitch™ · RolvKey™