The Zhitong Finance App learned that CMB International released a research report saying that on October 10, local time in the US, AMD (AMD.US) unveiled the new AI accelerator Instinct MI325X at the “NVDA.US” conference. The product is aimed at benchmarking the NVDA.US (NVDA.US) H200. The former is expected to be shipped in the fourth quarter of 2024. At this conference, AMD also unveiled the Mi350 series of chips that rival Nvidia's Blackwell series, but the series was mass-produced in 2H25 as soon as possible.
According to AMD CEO Lisa Su, the potential market size (TAM) for data center AI accelerators will grow at a compound growth rate of more than 60% and reach 500 billion US dollars in 2028, a sharp increase from previous forecasts (Lisa Su expected to reach 400 billion US dollars in 2027 in early 2024). Overall, the conference did not bring much surprise to investors.
At this meeting, I noticed the following three points: 1) Meta's Llama 405B model is fully operational on the MI300X, which means AMD has made good progress in Meta (META.US); 2) AMD did not mention Amazon (AMZN.US) during the customer presentation; 3) AMD did not disclose GPU sales targets for 2024 and 2025 and the current state of market supply and demand.
Highlights of AMD's new GPU: The MI325X accelerator uses a 5 nm process and is equipped with a 256GB HBM3e memory. The memory bandwidth reaches 6Tb/s. Its capacity and bandwidth are 1.8 and 1.3 times that of the Nvidia H200, respectively. According to AMD, the theoretical peak computing performance of the MI325X FP16 and FP8 will both reach 1.3 times that of the H200. Plans to achieve large-scale shipment of the MI325X in 4Q24 are progressing steadily. The product is expected to launch corresponding server solutions with various partners such as Dell (DELL.US), Eviden, Gigabyte, Lenovo, and ultra-microcomputers starting in 1Q25. In addition, AMD is also announcing a next-generation GPU MI350 accelerator based on the CDNA4 architecture and a 3nm process.
AMD said that the MI350X accelerator's computing performance (FP16 and FP8) will be 80% better than the MI325X. The product is equipped with 288GB HBM3e memory and supports 8Tb/s memory bandwidth. The MI350 series is expected to launch in 2H25.
AMD's new GPU highlights: In addition to releasing new GPU products, AMD also unveiled the fifth-generation EPYC “Turin” CPU based on the Zen 5 architecture at this conference. The performance of this CPU has been greatly improved compared to previous generation products, especially in data center applications. The Turin CPU increased the number of instructions per cycle (IPC) for enterprise and cloud computing workloads by 17%, while the number of instructions per cycle for performing high-performance computing and AI tasks increased significantly by 37%.
According to AMD, since launching the EPYC product line in 2018, AMD has increased its market share in the global server sector from 2% to 34%. AMD is also positioning the EPYC platform as an AI host CPU suitable for both AMD Instinct and Nvidia MGX/HGX platforms. These configurations can support up to 8 OAM MI300X or MI325X GPUs and provide outstanding performance benefits, including increasing AI inference performance by 20% and training workload capacity by 15%, positioning AMD as a key player in the AI CPU space and competing with Intel's Xeon series chips.
AMD's MI325X is a mid-term upgrade to the MI300X and is designed to compete with Nvidia's H200. However, since AMD's next MI350 is scheduled to be launched in 2H25, AMD will still lag behind Nvidia, given that the latter B200 will begin large-scale shipments in 4Q24. CMB International believes that Nvidia will continue to maintain its leading position in the GPU market, while AMD will continue to try to catch up. In terms of CPUs, AMD's 5th generation EPYC has made major breakthroughs and gained more market share in the server field. Its performance and cost efficiency are superior to Intel's Xeon 6 series.