CICC: Review of AMD "Advancing AI" Conference

According to the Zhitong Finance APP, CMB International released a research report stating that on October 10th local time in the United States, AMD (AMD.US) unveiled a new AI accelerator, Instinct MI325X, at the "Advancing AI" conference, aiming to compete with NVIDIA's (NVDA.US) H200. The former is expected to be shipped in the fourth quarter of 2024. At this conference, AMD also introduced the MI350 series chips to compete with NVIDIA's Blackwell series, but this series is expected to be mass-produced no earlier than 2H25. According to AMD CEO Lisa Su, the potential market size (TAM) of data center AI accelerators is expected to grow at a compound annual growth rate of over 60%, reaching $500 billion by 2028, a significant increase from previous forecasts (Lisa Su had previously predicted $400 billion by 2027 in early 2024). Overall, this conference did not bring many surprises to investors. Three points were noted at this conference: 1) Meta's Llama 405B model has been fully operational on the MI300X, indicating AMD's good progress in the Meta (META.US) area; 2) AMD did not mention Amazon (AMZN.US) in the customer showcase segment; 3) AMD did not disclose the GPU sales targets for 2024 and 2025, as well as the current market supply and demand status. Highlights of AMD's new GPU: The MI325X accelerator uses a 5nm process and is equipped with 256GB of HBM3e memory, with a memory bandwidth of 6TB/s, which is 1.8 times and 1.3 times the capacity and bandwidth of NVIDIA's H200, respectively. According to AMD, the theoretical peak performance of MI325X FP16 and FP8 calculations will both be 1.3 times that of H200. The plan for mass shipments of MI325X in 4Q24 is progressing steadily, and the product is expected to be launched with server solutions from partners such as Dell (DELL.US), Eviden, Gigabyte, Lenovo, and Supermicro starting in 1Q25. In addition, AMD also previewed the new generation GPU MI350 accelerator based on the CDNA 4 architecture and 3nm process. AMD stated that the computational performance (FP16 and FP8) of the MI350X accelerator will increase by 80% compared to MI325X. The product is equipped with 288GB of HBM3e memory and supports 8TB/s memory bandwidth. The MI350 series is expected to be launched in 2H25. Highlights of AMD's new GPU: 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 significantly improved compared to the previous generation, especially in the data center application field. The Turin CPU has increased the instructions per cycle (IPC) for enterprise and cloud computing workloads by 17%, and significantly increased the IPC for executing high-performance computing and AI tasks by 37% According to AMD, since the launch of the EPYC product line in 2018, AMD has expanded its market share in the global server field from 2% to 34%. AMD also positions the EPYC platform as suitable for both AMD Instinct and NVIDIA MGX/HGX platforms for AI host CPUs. These configurations can support up to 8 OAM MI300X or MI325X GPUs, providing outstanding performance advantages, including a 20% increase in AI inference performance and a 15% increase in training workload capacity, positioning AMD as a key player in the AI CPU field competing with Intel's Xeon series chips. AMD's MI325X is a mid-term upgrade of the MI300X, aiming to compete with NVIDIA's H200. However, as AMD's next-generation MI350 is planned to be launched in 2H25, AMD will still lag behind NVIDIA, considering that the latter's B200 will begin mass shipments in 4Q24. CICC believes that NVIDIA will continue to maintain its leading position in the GPU market, while AMD will continue to strive to catch up. In terms of CPUs, AMD's 5th generation EPYC has made significant breakthroughs, gaining more market share in the server field with performance and cost-effectiveness superior to Intel's Xeon 6 series.