AMD Launches AI Chip to Compete with NVIDIA; Market Projections Predict $500 Billion Growth in Four Years

AMD's Latest AI Computing Power Release and Market Response

On October 11, AMD held a launch event focused on artificial intelligence, unveiling several new products, with the MI325X computing chip taking center stage. Built on the CDNA 3 architecture, this chip serves as a significant mid-term upgrade, similar to the MI300X. The MI325X boasts an impressive 256GB of HBM3e memory and achieves a memory bandwidth of up to 6TB per second. AMD plans to begin production in the fourth quarter and anticipates supplying it through partnered server vendors early next year.

AMD aims for its accelerators to excel in content generation and inference within AI applications rather than merely handling large-scale data model training. The design of AMD's high bandwidth memory positions its chips to outperform Nvidia's in specific scenarios. While Nvidia's latest B200 chip features 192GB of HBM3e memory with an 8TB per second bandwidth, AMD’s Dr. Lisa Su highlighted that the MI325 outperformed Nvidia's H200 by 40% in the Llama 3.1 benchmark test.

Official documents indicate that the MI325 has superior specifications in peak theoretical FP16 and FP8 computing performance. AMD also announced plans to launch the MI350 series GPU next year, based on the CDNA 4 architecture. This series will include 288GB of HBM3e memory, leverage a 3nm process, and boost FP16 and FP8 performance by 80%, alongside a 35-fold increase in inference performance compared to CDNA 3. The MI355X GPU platform is expected to debut in the second half of next year, competing directly with Nvidia’s BlackWell architecture.

Dr. Su pointed out that the market for AI accelerators in data centers is projected to reach $500 billion by 2028, up from $45 billion in 2023. Despite this bright outlook, analysts estimate that Nvidia commands over 90% of the AI chip market, contributing to its 75% gross margin. Consequently, AMD's stock price saw a noticeable decline after the event, while Nvidia's shares soared nearly 180%.

Currently, AMD's primary revenue source in the data center sector remains CPU sales. In its recent financial report, AMD reported that data center sales doubled year-over-year to $2.8 billion, with AI chip sales contributing only $1 billion. AMD holds approximately 34% of the data center CPU market share. Additionally, the company introduced its fifth-generation EPYC "Turing" server CPUs, ranging from 8 to 192 cores, claiming that the EPYC 9965 outperforms Intel's Xeon 8592+ across multiple performance metrics.

To further reinforce its technological capabilities, AMD invited Kevin Salvadore, Vice President of Infrastructure and Engineering at Meta, who revealed that over 1.5 million EPYC CPUs have already been deployed. This collaboration highlights AMD's ongoing expansion and growth potential in the data center arena.