[Kim Ye-ji's Insight - Corporate Edition] Graphcore, which has proven its industry-leading performance in the IPU industry through the MLPerf 2.0 benchmark test for natural language processing, has entered the Korean AI semiconductor market in earnest by joining hands with ETRI, which is developing a new SW approach for high-efficiency AI computing.

Graphcore announced on the 6th that it has signed a partnership with the Electronics and Telecommunications Research Institute (ETRI) to improve AI computing and develop new software.

The size of AI models for natural language processing (NLP) has grown exponentially, and training massive AI models that can efficiently handle trillions of parameters has become increasingly challenging.It's getting worse.

The partnership between Graphcore and ETRI aims to develop a new software approach that improves AI computing performance, efficiency, and accessibility, while reducing the burden of ownership and dependence on single systems and technologies. Graphcore stated, "Rather than focusing solely on chips, we will focus on research into large-scale language models to explore efficient methods for natural language processing."

On the 7th, a media conference was held to announce the latest MLPerf results. In the closed-loop section of the MLPerf 2.0 benchmark test, Bow Pod16 demonstrated industry-leading performance by recording a training time 31% faster than NVIDIA's DGX-A100 640GB.
The 7-day event was attended by Graphcore Global Sales Vice President Febriz Moizan, Graphcore Korea Country Manager Kang Min-woo, and Large System Development Planning Vice President Phil Brown, and consisted of sessions such as Graphcore's △Korean business status △domestic and international customer cases △AI industry trends △Graphcore IPU system △introduction of the super-intelligent AI computer 'Good™ Computer'.

Graphcore's next-generation AI semiconductor intelligent processing unit, the 'IPU,' is already delivering performance benefits across a wide range of AI applications. Graphcore announced that it is developing next-generation IPU technology that will power the "Good Computer," a super-intelligent AI computer that surpasses brain-level parameters.

Graphcore, recognizing its status as a latecomer in Korea, announced plans to collaborate with startups and other partners, starting with AI developers, to provide optimization services. "TwigFarm" is a domestic example of a strategic collaboration that achieved 10x the performance and cost savings compared to existing GPUs.

Graphcore said, “We need an efficient approach with selective activation inspired by the routing function of the human brain,” and “Unlike GPUs, IPUs are chips that were born specifically for AI, and Graphcore’s ultimate design goal is to create AI that is close to the human brain.”

Regarding the packaging and other process technologies that will be used in the IPU, he said, “It is difficult to comment in detail on the process technologies that Good Computer Systems will adopt in the future,” but added, “We will further refine and develop packaging technologies, especially those related to 3D wafer integration, and bring about more innovation.”

Regarding ESG management issues such as the sustainability of data centers, he added, “Energy efficiency is also a key element that we are trying to improve through Good Computer,” and “Since the same effect can be achieved even with a smaller number of processes, this improved balance between memory and computers will enable us to achieve energy efficiency not only in large-scale models but also in small-scale models.”

Finally, Graphcore said, “Korea is an important AI market that surpasses the US and China. “Our government is emphasizing domestic AI semiconductors in its policies, and their importance will only grow further, including through collaboration with ETRI,” he said. “Machine learning (ML) is a key focus within the AI market, and this is the fundamental reason for developing the IPU.”