Arm has launched a pair of cores intended to bring acceleration for machine learning to its Cortex-M series of processors.
One is a complete Cortex-M, the M55, and the other is the U55, which is an add-on “micro neural processing unit” (microNPU) additional speed. According to the company, the M55 boosts inferencing performance up to 15-fold over the regular cores with a 5-fold improvement in more conventional DSP workloads. The main processor is the first in the M-series to be based on the Arm v8.1 architecture, which adds the Helium vector processing pipeline that is used to implement the machine-learning acceleration. As it supports the new architecture, the M55 also implements the custom-instruction support the company announced at Arm Techcon last year.
The Ethos-U55 provides an additional vector unit made up of hundreds of 8bit multiply-accumulate (MAC) units that, when used in parallel with the M55’s accelerator can give a 480x speedup compared to the existing Cortex-M processors. The accelerator takes up about 0.1mm2 of die space. With up to 50KB of onchip SRAM, the accelerator supports a number of memory optimizations that include various forms of compression and layer merging.
Tools support for the new cores include TensorFlow Lite Micro.