Graphcore boosts AI performance to exascale with wafer on wafer technology
Graphcore in the UK is the global lead customer for foundry TSMC for wafer-on-wafer technology to boost the performance of its AI chips.
The company has developed a new generation of its processor, using the second generation Colossus chips. The Bow IPU provides up to 40% more performance with 16% better power efficiency for real world AI applications, says the company. The BowPod 1024 rack unit with 1024 Bow processors will provide 350 PFLOPS of performance.
The company is using the wafer on wafer technology to build a large scale AI machine with 10 exaFLOPS of performance and 4Petabytes of memory and a bandwidth of 10 Pbyte/s, which will be comparable in complexity to the human brain.
“We are already working full speed on this today,” said Nigel Toon co-founder and CEO of Graphcore. “WoW and our silicon will make this possible.”
The Bow IPU has two wafers are bonded together to generate a new 3D die: one wafer for AI processing, which is architecturally compatible with the GC200 IPU processor with 1,472 independent IPU-Core tiles, capable of running more than 8,800 threads and 900MB of In-Processor Memory, and a second wafer with power delivery die.
Adding deep trench capacitors in the power delivery die next to the processing cores and memory enables the higher performance, says Simon Knowles, co-founder and chief architect at Graphcore.
Working closely with TSMC, Graphcore has qualified the entire technology including a number of ground breaking technologies in the Back Side Through Silicon Via (BTSV) and Wafer-on-Wafer (WoW) hybrid bonding.
“TSMC has worked closely with Graphcore as a leading customer for our breakthrough SoIC-WoW solution as their pioneering designs in cutting-edge parallel processing architectures make them an ideal match for our technology ,” said Paul de Bot, general manager of TSMC Europe.
“Graphcore has fully exploited the ability to add power delivery directly connected via our WoW technology to achieve a major step up in performance, and we look forward to working with them on further evolutions of this technology,” he said.
US cloud service provider, Cirrascale, is making Bow Pod systems available today to customers as part of its Graphcloud IPU bare metal service, while European cloud service provider G-Core Labs has announced that it will be launching Bow IPU cloud instances in Q2 2022.
One of the first customers will be the U.S Department of Energy’s (DOE) Pacific Northwest National Laboratory. (PNNL) for applications including cybersecurity and computational chemistry.
“For instance, we are pursuing applications in computational chemistry and cybersecurity applications. This year, Graphcore systems have allowed us to significantly reduce both training and inference times from days to hours for these applications. This speed up shows promise in helping us incorporate the tools of machine learning into our research mission in meaningful ways. We look forward to extending our collaboration with this newest generation technology,” said Sutanay Choudhury, deputy director of PNNL’s Computational and Theoretical Chemistry Institute.
Related Graphcore articles
