The E4, E6 and E8 Ensemble MCUs and fusion processors integrate the latest neural processing unit (NPU) from Arm, the Ethos-U85, with Alif’s high bandwidth memory and intelligent power management architecture to support NPU operators required by transformer networks to efficiently execute generative AI workloads.
Transformer networks enrich endpoint use cases by focusing on the most relevant parts of older portions of an acquired input sequence while rapidly processing in parallel the current acquired sequence to enable deeper contextualization. For localised use this results in an instant, accurate, and more satisfying experience when using multimodal AI in endpoint products focused on vision, voice, text, and sensor fusion without relying on the cloud.
The new Ensemble devices are based on a proven system architecture with large, fast on-chip tightly coupled memories to feed pairs of Arm Cortex-M55 CPUs coupled to Ethos-U55 NPUs, Alif’s proprietary aiPM granular power management, a hardware-isolated secure enclave, and multiple ways to stream image, sound, and sensor data into the device.
Ensemble devices can readily perform inferencing for neural networks using only its internal memory, and with external memory if required.
Commenting Paul Williamson, Senior Vice President and General Manager, IoT Line of Business at ARM, said, “This is the world’s first implementation of the Ethos-U85 in an endpoint AI MCU, and in Alif is enabling on-device generative AI, which is particularly valuable in voice and vision applications. OEMs can now embed AI into battery-powered applications, illustrating why edge AI is being built on the energy-efficient Arm compute platform.”
This second generation of Ensemble devices integrates a wider internal memory bus topology bringing abundant localised data bandwidth for NPU consumption, dual MIPI-CSI camera interfaces, a fast hardware image signal processor operating up to 200 frames per second, and a pair of faster interfaces to external memory each capable of 800MBytes/s. New broader options for SRAM data retention also give designers more flexibility to wake from sleep faster.
“Alif’s original Ensemble controllers in full production today remain at the forefront of the market for AI/ML performance, on-chip resources, energy efficiency, and small footprint. We specifically chose Arm’s standardised Ethos NPU IP because of its widely supported broad ecosystem in addition to superior performance and efficiency,” added Reza Kazerounian, President of Alif Semiconductor.
The second-generation Ensemble E4, E6 and E8 devices will be available for lead customers, while the first generation E1, E3, E5 and E7 Ensemble devices continue to be actively supported for new designs and are in volume production.
