N環球 
BrainChip Launches Sequora Processor IP to Run Generative AI on Battery-Powered Edge Devices
BrainChip's Sequora processor IP core, released Oct. 21, brings generative AI and LLMs to battery-powered edge devices like earbuds and sensor nodes, with constant memory requirements thanks to state-space model optimization. Now available for licensing.
Why it matters
The next battleground in edge AI is running without the cloud. BrainChip announced on Oct. 21 the release of its Sequora processor IP core, enabling generative AI and large language models to run directly on battery-powered edge devices such as earbuds and sensor nodes — from on-device LLM inference and voice-to-voice interaction to AI agents, all local, with no data sent to the cloud, improving privacy and latency.
Sequora complements the company's existing Akida neuromorphic processor IP: Akida handles streaming sensor data (vibration, audio, radar, raw RF IQ, biological signals) while Sequora takes on generative and sequential AI tasks.
Full report
The key technical bet is optimization for state-space models (SSMs), which maintain a fixed-size state to represent input history, keeping memory requirements constant regardless of sequence length — a fit for generative AI workloads on resource-constrained devices.
Sequora uses a unified multiply-accumulate (MAC) array for both signal processing and neural-network computation, eliminating the need for a separate DSP preprocessing stage, and its instruction-set architecture allows new algorithms to be deployed through software updates.
The IP is available now for licensing: process-portable, fully synthesizable RTL code including verification IP, testbench, synthesis scripts and timing constraints, plus an SDK with a static graph compiler, a C++ runtime library and a model zoo. Customers integrate it into their own system-on-chip and manufacture at the foundry and process node of their choice.
CEO Sean Hehir said: "Customers are not asking for just a signal classifier or a language model. They want both, in the same product, on the same battery."
The launch echoes an industry trend: the AI hardware bottleneck has moved from leading-edge logic down to mature nodes, and memory bandwidth and cost have become the leash on generative models. Running models at the edge with constant memory is a new route to lowering per-word cost.
💬 Comments
No comments yet. Be the first!