Australian chip company BrainChip Holdings has announced a new processor core designed to bring generative artificial intelligence, including large language models, to battery-powered edge devices that previously could only handle basic signal recognition.
Neuromorphic design targets low-power devices
The company said on October 14 that its new Sequora core IP extends its Akida neuromorphic platform so that devices can understand and respond to what they sense, rather than merely classifying signals, while keeping power consumption extremely low. Sequora uses a state-space architecture that can scale model sizes from a few million parameters up to three billion without increasing memory requirements, addressing the memory bandwidth constraints that have held back on-device generative AI.
BrainChip chief executive Sean Hehir said customers are looking for solutions that combine signal classification with language modelling on a single device without sending data to the cloud, pointing to privacy and latency advantages. The company said the hybrid design can process streaming sensor data while running generative tasks at the same time.
Licensable IP for custom chips
Sequora is being offered as licensable semiconductor intellectual property that can be integrated into custom system-on-chip designs, and comes with development tools for hardware makers. Frank Ferrante, chief executive of partner firm ForwardEdge ASIC, said there is strong demand for bringing generative AI into low-power custom silicon for critical applications.
BrainChip said the technology could be applied across aerospace, healthcare and consumer electronics, where devices need intelligence on board rather than depending on constant cloud connectivity.
Sources: MarketWire News, News Hub
Neuromorphic design targets low-power devices
The company said on October 14 that its new Sequora core IP extends its Akida neuromorphic platform so that devices can understand and respond to what they sense, rather than merely classifying signals, while keeping power consumption extremely low. Sequora uses a state-space architecture that can scale model sizes from a few million parameters up to three billion without increasing memory requirements, addressing the memory bandwidth constraints that have held back on-device generative AI.
BrainChip chief executive Sean Hehir said customers are looking for solutions that combine signal classification with language modelling on a single device without sending data to the cloud, pointing to privacy and latency advantages. The company said the hybrid design can process streaming sensor data while running generative tasks at the same time.
Licensable IP for custom chips
Sequora is being offered as licensable semiconductor intellectual property that can be integrated into custom system-on-chip designs, and comes with development tools for hardware makers. Frank Ferrante, chief executive of partner firm ForwardEdge ASIC, said there is strong demand for bringing generative AI into low-power custom silicon for critical applications.
BrainChip said the technology could be applied across aerospace, healthcare and consumer electronics, where devices need intelligence on board rather than depending on constant cloud connectivity.
Sources: MarketWire News, News Hub
