ARM turns over, Japanese semiconductor enters golden age again

Since AMD and Nvidia announced this year that they will accelerate the release of AI chips from "biennial updates" to "yearly updates", the competition among giants to catch up has become increasingly urgent. On October 10th local time, at the Advancing AI 2024 conference held in San Francisco, AMD officially launched the AMD Instinct MI325X, a GPU equipped with a new generation of AI chips.

According to AMD, the MI325X adopts the same CDNA 4 architecture as the previous generation MI300X, and its AI computing power (measured by half precision floating-point FP16) can reach up to 1.3 PFLOPS, which is the same as the MI300X. According to the official parameter comparison, MI325X directly competes with Nvidia's H200 GPU released in November last year, with computing performance about 1.3 times higher. At the same time, AMD followed up with Nvidia and for the first time enabled the industry's most advanced HBM3E high bandwidth memory on the MI325X, with memory bandwidth increased to 6TB/s and a total capacity of 256GB, both of which are higher than the MI300X.

According to AMD's official information, MI325X will be officially put into production in the fourth quarter of this year and delivered to customers starting in the first quarter of next year. Following the rhythm of "one iteration per year", AMD also revealed its future AI chip roadmap at the meeting. The next generation MI350 series will be launched next year, using a new CDNA 4 architecture that is different from MI300X and MI325X. The AI computing power under semi precision floating-point FP16 will reach 2.3PFLOPS. In addition to the new AI GPU, AMD has also released the fifth generation AMD EPYC server CPU processor specifically designed for enterprise customer data centers, as well as the Ryzen AI PRO 300 series processor for enterprise AI PCs.

AMD CEO Lisa Su predicted in her speech that the data center, AI, and accelerator markets will grow to $500 billion by 2028. This number is significantly higher than her prediction of "reaching $150 billion by 2027" from last year. She believes that generative AI plays a key role in this and requires significant investment in new infrastructure to support AI training and reasoning. Under such circumstances, the market is expected to grow at a rate of over 70% per year for the next four years.