A small AI chip startup called femtoAI just opened the doors of its developer community to everyone. On October 6, the San Bruno, California-based company announced that developers can now sign up for self-serve access to its full-stack AI silicon and compression platform, following what it describes as a successful beta program. One honesty note up front: everything below comes from femtoAI's own paid press release — no independent outlet has verified these claims yet.
What it is
femtoAI builds what it calls a full-stack AI silicon and compression platform. The centerpiece is the SPU — Sparse Processing Unit — a chip designed to exploit neural-network sparsity. The idea: in most neural networks, a large share of the weights don't meaningfully contribute to every computation. Skip the zeros and you burn less power and use less memory. That's the theory every efficient-inference startup is chasing.
What's now open to developers
Developers can create an account at developer.femto.ai and get access to: the SPU evaluation kit (order your own kit to test, develop, and prototype AI models on the actual SPU-001 chip); Sparsity Studio, a live tool that shows what sparsity does to your model — size, latency, and power — on your own network; the compiler and compressed models that make up the software side of the stack; model flexibility (bring your own model, use one of femtoAI's prebuilt models, or build from scratch, with what the company calls 'agent-friendly documentation'); and a community forum with open Q&A with femtoAI engineers and other developers.
The 10X claim — unverified
femtoAI says its patented sparsity and advanced model-compression techniques deliver 10X lower power draw and a 10X smaller memory footprint 'without sacrificing accuracy.' Per the company, it has demoed OpenAI Whisper running at a tenth of the memory. These are femtoAI's own numbers, from its own press release. Treat them as marketing until independent testing exists — 10X claims in chip land have a long history of melting under scrutiny.
Who's already using it
femtoAI names Legato and Marshall as beta customers — Legato makes smart glasses (Legato Frames), Marshall makes headphones and earbuds — plus an audio company called NewSound. Legato's CEO says integrating femtoAI's AI audio solutions helped achieve hearing performance in Legato Frames that was 'previously impossible in the form factor.' Again: customer quotes from the company's own release.
What's missing
No pricing is published for the evaluation kit or platform tiers, and per developer.femto.ai, access requires a free account plus a per-company NDA signed by an admin. No shipping timelines for ordered kits were given.
Verdict
This is an access opening, not a product launch — but it's worth noting because AI hardware is becoming a developer play, not just a hyperscaler play. If femtoAI's sparsity claims hold up under independent testing, cheap on-device AI inference gets a lot more interesting. Worth watching; verify before betting on it.
Comments
Post a Comment