Arm introduces next-generation GPU architecture

Recently, Arm demonstrated the next-generation graphics processing technology at the Arm DevSummit held online for shareholders and developers. The new GPU architecture is expected to be released in 2022. Arm said that compared with the current Mali-G710, the FP32 computing performance has doubled. Compared with the Mali-G76 in 2018, the FP32 ML performance has increased by 4.7 times, which is very obvious.

Ian Bratt, senior technical director of Arm’s ML business unit, demonstrated Arm’s latest ML performance for the first time but did not specify the power consumption of the new specification and the overall performance of the game. Compared with Mali-G78, Mali G-710 improves ML performance by 35% and graphics performance by 20% in ISO process node GPU configuration. The unnamed new GPU architecture will not be announced until next year, mainly for Chromebooks and high-end smartphones. For the time being, more details are lacking. Arm’s GPUs generally tend to be shipped the second year after the announcement.

Arm said that the new GPU architecture not only adds instructions and improves hardware IP, but also must provide software, tools, and libraries to achieve ML performance so that its processing power can be greatly improved. In order to speed up the advancement of machine learning tasks and other professional work, Arm is currently recruiting related development personnel to create its own project team.

According to The Register, Arm is focusing its GPU on the direction of neural network acceleration, which will increase the amount of calculation, require the correct type of calculation, and create tools to explore new neural network architectures. Nvidia also realizes that GPUs are very effective in processing AI tasks. The acquisition of Arm can further enhance its strength, can strengthen the operations of floating-point and integer operations, through software can use these operations to identify the pattern of the data set, and generate results faster.