Arm has added some of its machine-learning extensions to its latest Cortex-M series core, making it the smallest one in the company’s portfolio to date to feature AI acceleration.
The Cortex-M52 includes extensions based on the Helium architecture and is intended as an upgrade path for devices based on the M33 and M4 for developers who want to add dedicated neural-network support. According to Paul Williamson, general manager of the IoT business at Arm, the M52 speeds up typical machine-learning workloads up to 5.6 times compared to these earlier Cortex-M devices and 2.7 times for digital signal processing code.
In addition to the AI support, the core includes the security extensions form the v8.1M architecture, including TrustZone and the Pointer Authentication and Branch Target Identification (PACBTI) checks that were earlier added to recently designed v8 Cortex-A and Cortex-M cores. The PACBTI defences aim to head off the kind of pointer corruption often employed by hackers to take control of systems. The PAC instructions add authentication codes to the most significant bits of a pointer and are used by the processor to check that the address has not been altered to provide access to an area of memory not covered by that code. Similarly, the BTI additions check that the destination used by a branch instruction is a valid target and not one that has been modified at runtime.
Arm is making the core available on its cloud-based virtual hardware offering to allow remote prototyping of silicon designs and firmware.