Intel Unveils 'Gaudi3' to Compete with Nvidia... Builds AI Ecosystem with Naver (Comprehensive)

American semiconductor company Intel is concretizing an anti-NVIDIA alliance to increase its influence in the artificial intelligence (AI) era. It has decided to build a new AI semiconductor software ecosystem by opening an AI research center in collaboration with its key partner Naver. Intel unveiled its next-generation AI semiconductor 'Gaudi3' in the United States, which the company claims outperforms NVIDIA products and is based on an open ecosystem.

Pat Gelsinger, CEO of Intel, presenting at the 'Intel Vision 2024' event / Photo by Intel

원본보기 아이콘

On the 11th, Intel held a media briefing titled 'Intel Vision 2024' at the FKI Tower in Yeouido, Seoul, announcing these plans. The briefing was arranged to provide detailed information on the collaboration with Naver, which was revealed at the Intel Vision event held in Arizona, USA, on the 8th and 9th (local time). Lee Dongsu, Director (Ph.D.) of Naver Cloud, and Kwon Sejung, Leader of Naver Cloud, attended the meeting via video conference.

Director Lee said, "We plan to test using Intel Gaudi2 domestically," adding, "We will attach and evaluate various software to decide to what extent it will be applied internally." He also explained, "We intend to focus on exploring applicability in the large language model (LLM) field," and "The goal is very high as this collaboration aims to see whether several unique features or performances that have not emerged from existing AI semiconductors can be observed."

Naver Cloud will establish an AI joint research center (NICL) with Intel for this purpose. Centered on KAIST, it will collaborate with about 20 domestic research labs and startups, including Seoul National University, to explore ways to enhance Gaudi utilization. Through NICL, startups and universities will be provided with Gaudi-based IT infrastructure to conduct AI research. Various software development and industry-academia research projects based on Gaudi will also be operated.

Ha Jung-woo, Head of Naver Cloud AI Innovation Center, is presenting at 'Intel Vision 2024'. /
Photo by Intel

원본보기 아이콘

Director Lee said, "Gaudi is attractive because it offers better performance per watt," and added, "Since Intel has extensive experience operating data centers, we expect differentiation from the perspective of stable data center operation." The company also forecasted that after evaluations centered on Gaudi2, if the results are favorable, the possibility of adopting Intel's new product, the Gaudi3 accelerator, could increase.

Naver Cloud proposed a joint research approach to Intel to revitalize domestic AI research and strengthen the diversity of the AI chip ecosystem. Due to NVIDIA chips' high market share and dependency in the AI market, it has become difficult to receive products on time, and the burden of chip purchases has increased, worsening the AI resource environment for domestic startups and schools. This initiative aims to improve that situation.

Leader Kwon said, "Domestic startups and universities will directly use Gaudi2 and continue to create collaborations that share the results as open source," adding, "The goal is to make the results public and allow new users to enter the Gaudi2 ecosystem without fear by seeing those results."

Naver Cloud also announced plans to expand its ecosystem centered on its generative AI 'HyperCLOVA X' while leading this research. Kim Yuwon, CEO of Naver Cloud, emphasized, "For 25 years, Naver has provided users with alternative choices in a market dominated by global big tech companies, and diversity has been an important value pursued by Naver," adding, "In this context, securing diversity in the current AI chip ecosystem is very meaningful."

At Vision 2024 held in Phoenix, USA, Intel unveiled its latest AI semiconductor, Gaudi3. After showcasing a Gaudi3 prototype in New York last December, the product was officially released this time. Pat Gelsinger, Intel CEO, stated, "Gaudi3 is designed to optimize the operation of large-scale AI models," and expressed, "It demonstrated superior performance in benchmarks compared to H100 (NVIDIA AI semiconductor)."

According to Intel, Gaudi3 supports an average of 50% faster training time compared to the H100. Additionally, inference throughput is 50% higher than H100, and power efficiency is 40% better. Compared to its predecessor Gaudi2, it supports up to four times faster AI computing performance. Memory bandwidth has improved by 1.5 times, and networking bandwidth for large-scale system expansion has doubled.

Intel's newly released AI chip 'Gaudi3' / [Photo by Intel]

원본보기 아이콘

Notably, through software support based on an open community, companies can flexibly scale AI systems and applications, which differentiates it from NVIDIA products. NVIDIA has increased its market dominance through its software platform 'CUDA' that supports its AI semiconductors. Intel plans to accelerate the establishment of an open ecosystem based on Gaudi in partnership with companies such as Boshi, Red Hat, and VMware.

Na Seungju, Managing Director of Intel Korea, said at the briefing, "We now offer a better choice in the market from Gaudi2 to Gaudi3," adding, "It will be an opportunity to further improve performance and efficiency in terms of price, performance, and power consumption." He continued, "Based on Ethernet, an open industry standard, we are working closely with partners to expand into an open ecosystem without being dependent on any specific product."

Gaudi3 is manufactured using the 5-nanometer (nm; 1 nm = one billionth of a meter) process by Taiwanese foundry company TSMC. Like its predecessor, it features the third generation of high-bandwidth memory (HBM), HBM2E. The price has not been disclosed.

Intel will supply Gaudi3 to OEMs (Original Equipment Manufacturers) such as Dell Technologies, HPE, and Lenovo starting from the second quarter. From the third quarter, servers equipped with Gaudi3 are expected to be available in the market. Justin Hotard, Senior Vice President and General Manager of Intel's DCAI (Data Center and AI) group, explained, "Gaudi3 is a generative AI alternative with advantages in price-performance, system scalability, and availability."

© The Asia Business Daily(www.asiae.co.kr). All rights reserved.