It has strong AI processing, many independent IPU-Core tiles and capable of running more than 8,800 threads
The first Wafer-on-Wafer (WoW) processor named the Bow IPU by Graphcore aims to become an “ultra-intelligence AI supercomputer”.
The human brain is known to have billions of neurons that deliver fast computing. The purpose of the company is to build an AI computer that can beat this capability of the brain.
WoW bonds two wafers together to generate a new 3D die:
- The first wafer is for AI processing and is architecturally compatible with the GC200 IPU processor. It has 1,472 independent IPU-Core tiles, is capable of running more than 8,800 threads, and has 900MB of in-processor memory.
- The second wafer is a power delivery die and features deep trench capacitors that enable a large performance increase thanks to being located right next to the processing cores and memory.
Compared to its predecessors, the Bow IPU offers up to 40 per cent higher performance and 16 per cent better power efficiency for real-world AI applications. The flagship Bow Pod delivers more than 89 petaFLOPS of AI computing, which is increased by the superscale Bow POD to an incredible 350 petaFLOPS.
Other features include:
- Over 10 exaFLOPS of AI floating point compute
- Up to four petabytes of memory with a bandwidth of over 10 petabytes/second
- Support for AI model sizes of 500 trillion parameters
- 3D wafer on wafer logic stack
- Fully supported by Graphcore’s Poplar SDK
“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, GM of TSMC Europe.
By 2024, Graphcore expects to have delivered the first ultra-intelligence AI computer.