Graphcore UK launches new AI processor more involved than NVIDIA

Graphcore is a well-funded and ambitious British chip design company. It recently released the world’s most complex Colossus MK2 (also known as GC200 IPU) processor, claiming to have 59.4 billion transistors, and its performance is the same as that of the previous generation Colossus MK1 chip. Eight times. For comparison, Nvidia announced earlier this year A100 artificial intelligence chip, the number of transistors is 54 billion.

It is reported that each GC200 chip has 1472 independent computing cores/8832 independent parallel threads, all of which are supported by a built-in 900MB random access device (RAM).

Graphcore will equip its latest IPU model (M2000) with GC200 chips. It contains four GC200 chips in a pizza box-sized box, with a total computing power of 1 Peta-Flops.

At the same time as the announcement, artificial intelligence is having a major impact on the chip world. Because AI model training requires the use of highly parallel processors, in addition to spawning new market participants (Graphcore), it also encourages technology giants such as Google to accelerate the adoption of self-developed chips.

The company said that the new hardware fully supports plug and play, allowing customers to connect up to 64,000 IPU nodes together to achieve 16 Exa-Flops of computing power.

So far, NVIDIA has seized most of the AI ​​computing market. The company initially provided accelerated GPUs for video games, which has proven to be quite suitable for AI computing.

Even so, Graphcore is still trying to challenge Nvidia’s market dominance and has attracted funding and support from technology giants such as Microsoft and Dell.

Earlier this year, Graphcore announced that it had attracted US$150 million in research and development funds in a new round of financing, with a total valuation of US$1.95 billion.

As a company founded in 2012, the trend of deep learning was just beginning to emerge. However, the company’s biggest advantage is that its chips are fully considered for AI applications at the beginning of the design.

Moor Insights & Strategy analyst Karl Freund said in an interview with TheVerge that he was very impressed with Graphcore’s latest products. The software upgrade is particularly important for the correct use of the huge parallel processing capabilities required by AI.

Graphcore not only focuses on AI chips but also considers system-level integration. Because training a meaningful neural network cannot only rely on a single chip, but must be done on thousands or even tens of thousands of chip platforms, but the scalability makes Graphcore truly stand out.

(Via)