Amid a global investment boom in the artificial intelligence (AI) ecosystem, both corporations and academia in South Korea are struggling with the supply and expense issues of NVIDIA's graphics processing units (GPUs). In particular, with services and research and development becoming dependent on NVIDIA GPUs, major countries including the United States are grappling with alternatives. In South Korea as well, Intel, Naver, and KAIST have been collaborating to seek a way to decouple from NVIDIA, and they have recently achieved significant results.
On the 29th, Intel Korea conducted a 'Gaudi joint research results sharing meeting' with Naver Cloud and KAIST, revealing the research findings to date. Intel has promoted an industry-academic collaboration project involving Naver and major domestic universities (KAIST, Pohang University of Science and Technology POSTECH, Seoul National University). Verifications were conducted in various fields, including large language models (LLMs) and high bandwidth memory (HBM) design, resulting in cost-effective outcomes without NVIDIA GPUs.
Professor Kim Jong-ho of KAIST noted, 'This research aimed to foster growth in the AI industry by uniting fields previously separated, such as semiconductors, AI services, and academic research, and it began in earnest in August of last year.' He explained, 'From the perspective of researchers, this work based on Intel's Gaudi was successful to the extent that 15 papers emerged in just six months since the project started.'
'He added, 'This research achievement provides an opportunity to surpass institutions like MIT in the United States.' He said, 'With Intel's proactive support and Naver's software provision, we recorded numerous achievements, including leveraging AI for both enhanced learning and HBM design.' He continued, 'This type of collaboration model is rare globally, and given the chance, I want to try to change the NVIDIA-centered ecosystem with this type of AI research and development model.'
In terms of services, there have been empirical success cases as well. On this day, Naver is verifying the performance of Gaudi-based LLMs in a data center environment, claiming that it achieved inference performance up to 1.2 times faster than NVIDIA's A100, based on the open-source model 'LLaMA.' On this day, Lee Dong-soo, an executive at Naver Cloud, said, 'I hope this case results in more examples as a national agenda,' and added, 'Thanks to the efforts of Intel and domestic AI researchers, high-quality papers have emerged in just six months.'
Min-sung Cho, an executive at Intel Korea, remarked, 'The joint research case involving Intel, Naver, and KAIST has demonstrated how Intel's AI accelerator, Gaudi, can compete against NVIDIA.' He explained, 'Gaudi's advantage is not just its lower price compared to NVIDIA GPUs but also the verification of its performance through numerous experiments.' He added, 'Unlike NVIDIA's GPUs, which are dependent on specific networking, we can contribute to the expansion of the AI ecosystem by collaborating with a wider variety of companies.'
Intel is also strengthening its AI accelerator partnerships with IBM, Dell, and Microsoft (MS). Recently, Dell announced that its AI platform equipped with Intel's Gaudi3 showed excellent performance at a competitive price for generative AI workloads. It has been reported that Gaudi3 has demonstrated a performance-to-price ratio about 70% superior in inference throughput for the Llama3 80B model compared to the industry's leading GPU, NVIDIA H100.
Intel's Gaudi3 is gaining attention as a competitor to NVIDIA GPUs. According to a recent AI inference benchmark by the market research firm Signal65, Gaudi3 recorded 92% higher cost-effectiveness compared to competitors’ GPUs based on Meta's Llama-3.1-405B model. It also demonstrated a maximum advantage of 43% in tokens per second processed, proving its performance in AI inference workloads.