AMD Announced New MI325X Chip To Take On NVIDIA in AI Accelerator Race
AMD, which has always been overshadowed by NVIDIA, just held another AI-themed press conference on Thursday, launching a range of new products, including the MI325X computing chip. However, amid lukewarm market enthusiasm, AMD's stock price also took a noticeable plunge.
As the most market-focused product, the MI325X, like the previously launched MI300X, is based on the CDNA 3 architecture and has a similar basic design. Therefore, the MI325X can be seen more as a mid-life upgrade, featuring 256GB of HBM3e memory with a memory bandwidth of up to 6TB/s. The company expects this chip to start production in the fourth quarter and will be supplied by cooperating server manufacturers in the first quarter of next year.
In AMD's positioning, the company's AI accelerators are more competitive in use cases for creating AI model content or performing inference, rather than training models by processing massive amounts of data. Part of the reason is that AMD has stacked more high-bandwidth memory on the chip, allowing it to perform better than some NVIDIA chips. Comparatively, NVIDIA has equipped its latest B200 chip with 192GB of HBM3e memory, which is two B100s each connected to four 24GB memory chips, but the memory bandwidth can reach 8TB/s.
AMD CEO Dr. Lisa Su emphasized at the press conference that the MI325 delivers up to 40% higher performance than NVIDIA's H200 when running Llama 3.1.
According to official documents, the MI325, which has parameter advantages, can provide 1.3 times the peak theoretical FP16 (16-bit floating-point number) and FP8 computing performance compared with the H200.
Compared with the MI325X, AMD has also painted a big cake for the market - the company will launch the MI350 series of GPUs based on the CDNA 4 architecture next year. In addition to the further increase in HBM3e memory scale to 288GB and the process technology being upgraded to 3nm, the performance improvement is also astonishing. For example, the FP16 and FP8 performance are 80% higher than the just released MI325X. The company even stated that compared with CDNA 3 accelerators, the inference performance of the MI350 series will be increased by 35 times.
AMD expects the platform equipped with the MI355X GPU to be launched in the second half of next year, directly competing with NVIDIA's BlackWell architecture products.
Dr. Lisa Su also stated on Thursday that the market for data center artificial intelligence accelerators will grow to $500 billion by 2028, up from $45 billion in 2023. In previous statements, she had expected this market to reach $400 billion by 2027.
It is worth mentioning that most industry analyses believe that NVIDIA's market share in the AI chip market can reach more than 90%, which is also the reason why the chip leader can enjoy a gross margin of 75%. Based on the same consideration, the difference in stock performance between the two parties is also very large - after the press conference today, AMD's (red line) year-to-date increase narrowed back to within 20%, while NVIDIA's (green line) increase is close to 180%.
For AMD, the majority of the current data center business still comes from CPU sales. In practical use cases, GPUs also need to be used in conjunction with CPUs.
In the financial report for the quarter of June, AMD's data center sales doubled year-on-year to $2.8 billion, but AI chips only accounted for $1 billion. The company stated that its market share in the data center CPU market is about 34%, lower than Intel's Xeon series chips.
As a challenger in the data center CPU field, AMD also launched the fifth-generation EPYC Toppling series server CPU on Thursday, with specifications ranging from the 8-core 9015 ($527) to the highest 192-core 9965 ($14,831). AMD emphasized that the performance of EPYC 9965 is several times stronger than Intel's flagship server CPU Xeon 8592+.
At the press conference, AMD invited Meta's Vice President of Infrastructure and Engineering, Kevin Salvadore, to support the stage, who revealed that the company has deployed more than 1.5 million EPYC CPUs