Intel, Naver to open joint AI research institute

(Right) Ha Jung-woo, Head of AI Innovation at Naver Cloud, speaks with Intel CEO Pat Gelsinger during the Intel Vision 2024 keynote speech session, in Phoenix, Tuesday. [Courtesy of Intel Corp.]

U.S. chip giant Intel Corp. will collaborate with South Korea’s Naver Corp. to open an artificial intelligence (AI) semiconductor research institute. The move is part of an effort to form a united front to counter Nvidia Corp.’s dominance in the AI chip market.

“Naver is an outstanding company that has built tremendous AI models in Asia,” Intel chief executive Pat Gelsinger said during the Intel Vision 2024 event in Phoenix, Arizona, the United States, on Tuesday local time.

Naver chief executive Choi Soo-yeon responded that the company “is very excited to work together with Intel.”

The collaboration between Intel and Naver is aimed at expanding the open platform software ecosystem based on Intel’s AI accelerator, Gaudi. To achieve this, both companies are investing jointly to open an AI chip research institute at Naver’s headquarters in Bundang, Gyeonggi Province.

[Graphics by Song Ji-yoon and Lee Eun-joo]

Intel’s AI specialist division Habana Labs will participate in the project alongside Naver’s semiconductor researchers, and Lee Dong-soo, director of hyperscale AI at Naver Cloud, will direct the institute.

The reason for the duo’s collaboration lies in Nvidia’s monopoly, with the company’s CUDA currently monopolizing the parallel processing software market necessary for AI training and inference. With over 80 percent market share in the graphics processing unit (GPU) market, Nvidia naturally dominates the AI accelerator development platform as well.

The problem lies in GPU procurement, as Nvidia’s H100, priced at about $40,000, reportedly takes more than 50 weeks from order to delivery and A100, the low-cost GPU, has been discontinued. This makes it difficult for Naver to update its super AI HyperCLOVA X.

Although entering the market is challenging for Intel, it unveiled the AI accelerator Gaudi3, which significantly reduces AI model training time, on Tuesday. Gaudi3 boasts training speeds twice as fast as Nvidia’s flagship GPU, H100 and Intel highlighted that based on tests using Meta’s LaMa2 model alone, Gaudi3 exhibited 50 percent higher inference throughout and 40 percent better power efficiency than H100. The company also said that Gaudi3 has four times the AI computing power in BF16 compared to previous Intel products, and 1.5 times the memory bandwidth.