Arm 2022 GPU Architecture to Deliver 5X ML Workload Performance
October 22, 2021
Arm has announced an unnamed GPU architecture will be released next year that delivers a 4.7x performance increase in FP32 workload performance.
Speaking from the company’s virtual Arm DevSummit this week, Ian Bratt, a Fellow and Senior Director of Technology for ML at Arm, unsurprisingly said that the new GPU architecture will cater to AI and ML workloads.
The expected 4.7X increase in the 2022 architecture compares to the Mali-G76, which was released in 2018 and is based on the previous-generation Bifrost architecture. Compared to the Mali G710 – the company’s third GPU built on the current-generation Valhall architecture – the 2022 GPU is projected to offer double the FP32 machine learning performance.
Bratt said that precise performance-per-watt metrics are still unknown, and may not available for some time given Arm cores typically don’t ship in production silicon for a year after release. For context, the G710 came out of the gate with a 35% improvement in ML workloads and 20% higher graphics performance in ISO-process node GPU configuration compared to its G78 predecessor.
The Arm Mali G710 is slated to ship in laptop, tablet, and smartphone sockets this year.
According to Bratt, Arm expects to continue developing GPU architectures on a three-year release cadence to address the expanding demand for ML performance. They hope to achieve this through better per-core performance, support for increasing numbers of cores, and software development infrastructure for advanced neural networks that mirror human cognition.
"We actually don't have the luxury of millions of years of evolution, so we need to develop tools to kind of short circuit that evolution and enable quick exploration of neural network architectures,” he said. "It's more than just adding instructions and improving hardware IP, we also have to provide the software, the tools, the libraries to enable that ML performance."