Graphcore launches Mk2 IPU systems as direct competitor to Nvidia A100 systems

[agustincosta]

Summary

 

Graphcore, a UK-based chip design company, launched its second iteration of their Intelligence Processing Unit (IPU) for AI acceleration, which they claim is the world's most complex chip: the Colossus MK2 or GC200 IPU. The processor has 59.4 billion transistors and offers an eight times performance increase from the company’s Colossus MK1, says Graphcore. It boasts more than the 54 billion transistors found in Nvidia’s A100, announced earlier this year.

 

Graphcore will be making the GC200 available via its new IPU Machine, the M2000, which contains four GC200 chips in a single chassis and which delivers 1 petaflop of total compute. The firm says its new hardware is “completely plug-and-play” and that customers will be able to connect up to 64,000 IPUs together for a total of 16 exaflops of computing power.

 

 

Quotes

Quote

 With 59.4Bn transistors, and built using the very latest TSMC 7nm process, the Colossus MK2 GC200 IPU is the world's most sophisticated processor. Each MK2 IPU has 1472 powerful processor cores, running nearly 9,000 independent parallel program threads. Each IPU holds an unprecedented 900MB In-Processor-Memory™ with 250 TeraFlops of AI compute at FP16.16 and FP16.SR (stochastic rounding). The GC200 supports much more FP32 compute than any other processor.

 

The IPU-M2000 is Graphcore's new breakthrough IPU system built with our second generation IPU processors for the most demanding machine intelligence workloads. Our advanced architecture delivers 1 PetaFlop of AI compute and up to 450GB super-fast Exchange-Memory™. The IPU-Machine features our new ultra-low latency IPU-Fabric™ to build scale-out IPU-POD™ data centre solutions connecting up to 64,000 IPUs.

 

My thoughts

This new product could be very interesting as a competitor for Nvidia, which has dominated the market so far in AI compute. There seems to be a lot of work put into this design, both in the electronics and software side, and it could be a gamechanger for its price to performance numbers against other competitors.

 

Sources

https://www.graphcore.ai/products/ipu

https://www.youtube.com/watch?time_continue=1&v=_zvU0uwIafQ

https://www.theverge.com/2020/7/15/21324103/ai-chips-nvidia-graphcore-colossus-mk2-gc200-ipu

https://www.forbes.com/sites/moorinsights/2020/07/15/could-graphcores-second-chip-challenge-nvidia/#425ce1c24e5c

[tenchi71]

But can it run Crysis?

[williamcll]

I'll wait for the benchmarks, if it ever appears. 

[Results45]

Not holding my breath. Fujitsu, Huawei, and Intel also claim to have custom solutions that can process massive GPU workloads efficiently.

 

https://www.nextplatform.com/2019/11/13/a64fx-arm-chip-gets-a-big-push-from-cray/