Arm gets edgier with faster machine-learning accelerator to slot alongside Cortex, Neoverse CPU cores
Ethos-U65 sounds like an industrial techno band from the 1990s but, well, it isn't
Arm this week launched its Ethos-U65 AI accelerator for running workloads on edge devices and other gear.
The design is described as a microNPU, where the NPU stands for neural processing unit; it's supposed to be used alongside Cortex-M, Cortex-A, and Neoverse-series CPU cores, which are aimed primarily at microcontrollers, smartphones and laptops, and servers, respectively. The CPU cores offload AI math onto the microNPU, and they're all packaged up into a system-on-chip.
A quick look at its specs [PDF] reveals that the Ethos-U65 is optimized for running convolutional neural networks, recurrent neural networks, and long-short term memory networks. These are typically used in computer vision and natural language processing applications.
Thus the technology can be used for things like real-time face, object, and speech recognition, for instance. Arm pitched the Ethos-U65 as bringing AI to the latest IT trend: the network edge. That means we're likely to see the U65 used in things like smart-home devices and cameras, personal digital assistants, and controllers that manageme equipment and sensors.
Arm hasn't given up on self-driving car brains – its new Cortex-A78AE is going into Nvidia's Orin chip, for a startREAD MORE
The Ethos-U65 takes up about 0.6mm2 of die space when using, say, a 16nm fabrication node, supports INT8 and INT16 precision – which is about all you need for modest AI inference – and hits a maximum performance of one trillion operations per second (1 TOPS at 16nm).
It has an internal SRAM of 55 to 104KB, and can process 256 or 512 multiply-and-accumulate operations per clock cycle. The latter is essential for performing vector math, which is the bedrock of today's neural network algorithms. You can tap into the Ethos-U65 using Google’s TensorFlow Lite Micro framework, and it can run popular off-the-shelf models, such as Google’s Inception V3 for object recognition, or Facebook’s Wav2Letter for speech recognition.
The Ethos-U65, which is available for licensing, follows last year's U55, and is supposed to outperform its predecessor, as you might expect. That means stuff using chips that incorporate the U65 should work better than previous generations, ultimately.
“Adding to the success of the Ethos-U55, and maintaining a focus on power efficiency, the Ethos-U65 extends its applicability to Arm Cortex-A and Neoverse based systems,” Arm's Tanuj Arora said.
“This enables new capabilities in devices such as high-resolution smart cameras, smart home solutions, and even infrastructure applications such as bandwidth and power management subsystems.” ®