First ARMv9 processor for edge AI is ‘fundamental shift’
Cette publication existe aussi en Français
ARM has developed its first processor core for edge AI and embedded applications using the ARMv9 instruction set. This is a fundamental shift for embedded developers who would previously have used a high end microcontroller, says Paul Williamson, general manager of the
The move to a microprocessor rather than a microcontroller is all about the memory management for AI frameworks. UP to four ARM Cortex-A320 cores can be clustered together with a U85 neural processing unit (NPU) for an embedded processor.
“This isn’t just an incremental step forward, it’s a fundamental shift and we believe it will drive the edge AI forward for many years to come,” he said. “We developed this from the ground up specifically for edge AI.”
- Paul Williamson interview: ARM moves to support transformer AI models in embedded applications
- Low power M55 microcontrollers with U85 for edge AI
- Renesas ships M85 microcontroller
The A320 will be suitable for embedded developers as it can also run existing real time operating systems such as FreeRTOs and Zephyr, as well as high level Linux and Android operating systems as well as tuned AI models of over 1bn parameters such as large language models (LLMs) or a selection of smaller models for agentic AI.
“We can only realise the potential of AI if we move it to the devices around us. A few years ago the AI workloads were simpler, with noise reduction or anomaly detection but now there are much more complex so we are seeing the need for higher performance,” said Williamson.
“The ability to run an RTOS provides a development path for microcontrollers with increased memory flexibility and we believe the A320 offers the maximum flexibility,” he said. “I think you will see completely new configurations of what was previously done in a microcontroller.”
This does not mark a move away from microcontrollers, he says, even though the Cortex-M85 has already been paired with the U85 in the RA8E microcontroller families from Renesas Electronics launched back in October 2023 and dual M55 cores paired with the U85 in the E series from Alif Semiconductor.
“Microcontrollers are optimised for power and the M85 is still early into the market and we are really pleased with the platforms that are emerging. But there are some things that M class doesn’t handle very well, particularly large memory systems and AI, and those two trends make an interesting opportunity,” said Williamson.
“The ARMv9 enhancements improve machine learning performance with 8 times the performance of the Cortex-M85 with U85 for transformer networks,” he said. “We have updated the driver for the U85 to be driven directly form the A320 without the need for the M85 microcontroller in there are well.
He sees the combination of the M85 and U85 as ‘endpoint AI’ vs edge AI with the A320 and U85. ARM has signed a deal with Amazon Web Services (AWS) to support the chip in its Greegrass IoT service. “Systems with better memory performance are becoming more necessary and the A class is more flexible in handling multiple tiers of memory latency,” he said.
The ARMv9 architecture also provides more security with memory tagging extensions, pointer authentication and level 2 secure exception in the TrustZone security block to support more secure containers which are often used for the AI frameworks.
The A320 is the lowest tier of the V9 family, which extends to the A520 and A725 and up to the high end X925 core and Williamson sees the core being used in chips that range from consumer wearables through precision agriculture and smart manufacturing to autonomous vehicles as well as in baseboard management in servers and infrastructure
“There’s quite a wide range in potential performance,” he said “We expect this on 22nm ultra low power at 1.2GHz which is very different to 5nm FINFET at 2.3GHz.”
The ARM Cortex-A320 is in development with partners with silicon next year, while a fast model of the Cortex-A320 is currently available to partners for ARMv9 software development. A cycle-accurate fixed virtual platform (FVP) of the combined Cortex-A320 and Ethos-U85 will be available later this year.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
