SureCore IP slashes AI chip power at KU Leuven
Cette publication existe aussi en Français
KU Leuven in Belgium is using IP from SureCore to reduce dynamic power in an AI chip by 40%.
The PowerMiser IP was used in a 16nm FinFET process and can reduce dynamic power by up to 50% and static/leakage power by up to 20% compared to foundry and other SRAM solutions, with savings across the full process, voltage and temperature range. SureCore in Sheffield, UK, is planning a 7nm variant of the technology.
“People forget that the initial drivers for the 16nm node were mobile and HPC solutions, and hence most of the IP developed for this node was optimised for performance not power,” said said Paul Wells, CEO at sureCore. “Today 16nm could almost be considered to be a mature node with many millions of devices in the field. Forward-thinking application developers are now looking to exploit this node’s improved density, leakage and power characteristics, especially for wearables, medical and Edge-AI devices. This is where our PowerMiser SRAM can bring huge benefits by enabling challenging power budgets to be delivered.”
“We licensed sureCore’s PowerMiser IP because we wanted to create a novel neural processing accelerator chip for AI applications. The chip has very high computational processing needs, and, of course, such devices naturally also have significant power consumption characteristics,” said Professor Wim Dehaene at KU Leuven, who is working in the MICAS research division of KU Leuven’s electrical engineering department.
MICAS last year developed a digital in-memory computing (DIMC) for vision recognition which uses substantial amounts of SRAM memory. The HUNBN 16 nm FinFET chip demonstrates a complete IMC based accelerator. It shows that DIMC is an energy-friendly alternative for classic accelerators, especially when the model is moderately sized and can fit at once on the chip. This gave a neural network chip with a best efficiency of 24 TOPS/W at 4bit quantisation and 30 frames/s performance.
www.micas.be; www.surecore.com
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
