FFN speedups for edge inference
Threads from @rolveitrem show FFN (feed‑forward network) acceleration techniques claiming 5–300x speedups and 80–99% energy reductions on commodity hardware (Xeon laptops, TPUs), enabling edge‑deployable model inference without GPUs shared and expanded.
ROLV LLC — led by founder Rolv E. Heggenhougen — published a company press release on Oct. 1, 2025 announcing the rolvsparse library and related benchmark claims. (pr.com) Benchmarks published on the rolv site show a Llama‑4 Maverick MoE expert FFN test where cuBLAS throughput of 369k tokens/s was increased to 7.66M tokens/s with rolvsparse, and the report lists energy numbers of 232.32 J (cuBLAS) versus 42.97 J (rolv) for the same run. (rolv.ai) The company says its test artifacts were independently validated by the University of Miami, with output hashes reproduced across NVIDIA, AMD, Intel Xeon, and Apple M4 Pro platforms and a Zenodo record cited for cross‑platform verification in December 2025. (rolv.ai) ROLV publishes a public verification kit and benchmark PDF on GitHub (rolv‑verifier) that the firm invites third parties to run for reproducibility on local hardware. (github.com) The Oct. 1, 2025 announcement also cites a fast‑track parent patent plus four Continuation‑in‑Part filings backing the ROLV Library, and the company contrasts cost by claiming a $2,000 dual‑Xeon server with rolvsparse can match or beat a $40,000 NVIDIA B200 at ≥80% sparsity. (pr.com) Community discussion and independent checks have appeared on Hacker News and other forums, where contributors have posted reproducibility questions and run independent microbenchmarks against the provided artifacts. (news.ycombinator.com)
