Intel's surprise attack on Nvidia

Author | Zhou Zhiyu

In the fiercely competitive AI arena, no one wants to see NVIDIA dominate alone.

On April 10th, Beijing time, Intel released a combination of hardware and software products including the AI chip Gaudi 3 and the all-new brand of Xeon 6 processors, and announced the launch of the next-generation Core Ultra processor family codenamed "LunarLake" later this year. In the enterprise and consumer markets, Intel is accelerating the launch of new AI products.

Gaudi 3 was the star product of this conference, as Intel aims to compete with NVIDIA and win over AI users from NVIDIA.

On the day of the conference, NVIDIA's stock price fell nearly 5% at one point, ultimately closing down 2.04%, while Intel rose 0.92%, reflecting to some extent investors' approval of Intel's latest achievements.

Intel CEO Gelsinger was also quite satisfied with Gaudi 3. At the conference, Gelsinger danced to the music, showcasing Gaudi 3 to the world.

Intel revealed that compared to NVIDIA's H100 GPU, Gaudi 3 has a 70% improvement in training performance, a 50% improvement in inference performance, a 40% improvement in energy efficiency, and costs only "a small fraction" of the H100. Compared to NVIDIA's latest flagship chip, the H200, Gaudi 3 also outperforms it by 30% in inference performance. This chip is manufactured using TSMC's 5-nanometer process.

In addition, Intel also announced a refresh of the Intel Xeon 6 brand, with the introduction of Ecores-equipped Xeon 6 processors in the second quarter of this year, followed by the release of P-cores-equipped Xeon 6 processors with higher AI performance. The Intel Xeon 6 processor with P-cores will be able to run the Llama-2 model with 700 billion parameters.

Maintaining a performance lead in inference performance while focusing on energy efficiency is Intel's confidence in challenging NVIDIA head-on. The core of this lies in the shift of competition focus in the generative AI field from last year's "battle of hundred models" to the battle of applications.

Last year, the core of competition among AI companies was in training large general models, where companies were not too concerned about costs and power consumption. However, this year, most models on the market are based on open source, with little difference in performance and training data, making it difficult to differentiate and profit at the training end. The focus of AI companies has gradually shifted to how to use large models to solve specific business problems, enhance product competitiveness, and other companies also hope to find suitable large models to integrate into production or business processes to create value.

Preferences for AI chip performance and data center construction have also shifted from training capabilities to inference capabilities.

Within the industry, some companies are also exploring the possibility of using CPUs for large model inference, including tech companies like Qualcomm and Microsoft who have noticed this trend and introduced products such as Qualcomm's Oryon CPU that can run 13 billion parameter large models on PC terminals, and Microsoft's self-developed AI chips Azure Maia 100 and Cobalt 100 CPU Vice President of Marketing Group and General Manager of Data Center Sales in China, Eddie Wu, pointed out that power consumption is very important for inference because it requires large-scale deployment. It is necessary to consider the cost, power consumption, and overall operational aspects of the intelligent computing center.

Wu mentioned that in the process of building new data centers, Intel also found that last year, resources were primarily focused on intelligent computing data centers. However, from the second half of last year to this year, the construction focus has shifted to maintaining high-speed growth in intelligent computing data centers while general data centers are also returning to their previous construction pace.

Li Liangya, General Manager of Cloud and Industry Solutions Department at Intel China, also stated that considerations for green and energy-saving technologies may bring new opportunities for the future development of data centers. When enterprises need to address monetization issues after the hype of large models fades, they need to think about the economic applicability and the most suitable solutions.

All of this brings new development opportunities for Intel. Intel expects that enterprise investment in generative AI will reach $40 billion this year and will reach $151 billion by 2027.

Over the past decade, due to factors such as the continuous postponement of manufacturing process node introductions, Intel has been surpassed by companies like Samsung in terms of revenue, followed by TSMC, AMD, and NVIDIA in terms of market value.

The once semiconductor giant has faded in the AI era. Since the beginning of this year, Intel's stock price has fallen by 24.89%, while NVIDIA has risen by 82.99% and AMD has risen by 15.66% during the same period.

Intel wants to regain the lost decade and has adjusted its business focus towards the AI wave. The wafer foundry business operates independently, and products are geared towards the AI wave, which of course brings financial pain. However, in the fourth quarter of last year, Intel's revenue and profit exceeded external expectations. Krzanich stated that the fourth quarter marked further progress in Intel's transformation over the past year, and the company will continue to focus on wafer foundry operations and popularize AI this year.

Krzanich predicts that the semiconductor market will reach $1 trillion by 2030, with AI being the main driving force.

In the coming years, the AI wave will continue, and various technology companies will continue to compete on the software and hardware fronts to become the king of the AI era. Intel is also attempting a turnaround to return to its peak in the new era