femtoAI
femtoAI builds sparse AI accelerators and software for ultra-low-power edge inference.
- 24
- Seed
femtoAI builds sparse AI accelerators and software for ultra-low-power edge inference.
femtoAI differentiates with a full-stack sparse inference approach that co-designs models and silicon for dual sparsity, aiming for large capacity gains and large energy reductions on extreme-edge workloads. The SPU architecture is described as tileable and programmable across CNN, RNN, transformer, and custom models rather than a fixed ASIC for one network family.
That combination of software tooling and first-party accelerator silicon helps customers who need always-on audio or sensing functions where GPU-class power budgets are unavailable.
Source: femto.ai
Demand for on-device AI continues to rise as product teams push intelligent features into wearables, appliances, hearables, robotics, and industrial sensors that cannot rely on continuous cloud connectivity. Power, latency, and cost constraints favor specialized sparse accelerators over general-purpose GPUs for these always-on use cases.
Vendors that pair model compression with purpose-built silicon are positioned to capture design wins where memory bandwidth and milliwatt budgets dominate system architecture.
Source: femto.ai
femtoAI sells an embedded AI inference platform spanning ready-to-deploy sparse models, an SDK, and the Sparse Processing Unit accelerator family. SPU-001 is positioned as a low-power AI co-processor with on-chip memory and interfaces suited to speech, audio, and related one-dimensional workloads.
Complementary evaluation boards and kits let OEMs integrate the silicon with host processors and hearing-aid or sensor platforms before volume deployment.
Source: femto.ai