show image

AI chipmaker Graphcore releases “most complex processor ever made”

The British chipmaker Graphcore has won praise from analysts after releasing the second generation of its intelligence processing unit (IPU) chip, the GC200, which it claims is the most “complex processor ever made”.

The chip powers a new IPU machine, the M2000, and represents a different approach to AI hardware from the graphics processing units (GPUs) and central processing units (CPUs) produced by Nvidia and Intel respectively.

The M2000 contains four IPU chips, delivering “a plug-and-play petaflop of compute in a unit the size of a pizza box”, according to Graphcore. The IPU machines can operate as standalone units or be packaged together in data centre pods containing up to 64,000 GC200 chips.

“Graphcore’s intelligent processing unit (IPU) is more brain-like than Nvidia’s GPUs or Intel’s CPUs with which it is competing primarily in the AI arena,” said Mike Orme, a thematic research consultant at GlobalData, which is part of the same group as NS Tech.

“The new IPUs are based on wafer-scale chips each containing 1,472 processor cores each with its own superfast flash memory modelled on the brain’s neuron-synapse structure,” Orme added. “This eliminates the bottleneck of the classical computing model, used by Nvidia and Intel, of moving data on to the chip for processing and off again, which is waste of time and energy. With the IPU, it’s all on the chip.”

The race to develop ever smaller, faster and cheaper AI chips is one of the most fiercely contested in the tech industry. According to some estimates, the AI chip market will be worth $91.2bn by 2025. Graphcore raised an addition $150m in May at valuation worth nearly $2bn.

The company operates in a range of sectors from automotive to healthcare and boasts a client list including Oxford University, the US hedge fund Citadel and Microsoft, which is using the chips to power AI applications hosted in its Azure cloud computing service. The Microsoft deal was announced last November and gives Azure users the ability to access Graphcore’s hardware from any location.

Karl Freund, a senior analyst at Moor Insights who has written a report on the new chips that has been distributed by Graphcore, said its latest product launch put the Bristol-based business “first in line to challenge NVIDIA for datacenter AI”.