AMD (AMD) Reveals MI325X Chip Amid AI Expansion Efforts

Author's Avatar
2 hours ago
Article's Main Image

AMD, often overshadowed by NVIDIA in the AI computing sector, unveiled a new suite of products, including the MI325X chip, at a recent AI-focused event. Despite the launch, AMD's stock faced a notable drop. The MI325X, similar to the MI300X, is based on the CDNA 3 architecture and is considered a mid-term upgrade. With 256GB of HBM3e memory, it offers a maximum memory bandwidth of 6TB/s. Production is expected to start in the fourth quarter, with supply through server manufacturers in the first quarter of next year.

AMD positions its AI accelerators as more competitive in AI model creation and inference applications rather than in processing large datasets. This is partly due to AMD's use of higher bandwidth memory, which can outperform some of NVIDIA's chips. For instance, NVIDIA's latest B200 chip features 192GB of HBM3e memory, with a bandwidth reaching 8TB/s.

AMD's CEO, Lisa Su, highlighted that the MI325 can deliver up to 40% higher performance than NVIDIA's H200 when running Llama 3.1. AMD's official documents indicate that the MI325 offers 1.3 times the peak theoretical FP16 and FP8 computing performance compared to the H200.

Looking ahead, AMD plans to release the MI350 series GPUs based on the CDNA 4 architecture next year. These chips will feature an expanded 288GB HBM3e memory and a 3nm process, with performance improvements of 80% over the MI325. The MI350 series is expected to provide 35 times the inference performance of CDNA 3 accelerators.

AMD predicts that the market for AI accelerators in data centers will grow to $500 billion by 2028, up from $45 billion in 2023. Previously, Su projected this market to reach $400 billion by 2027.

NVIDIA currently holds over 90% market share in AI chips, which results in a 75% profit margin. As a result, there's a stark contrast in stock performance: post-event, AMD's year-to-date gains have slipped below 20%, whereas NVIDIA's have surged about 180%.

In terms of CPUs for data centers, AMD's business remains heavily reliant on CPU sales. AMD's data center sales doubled to $2.8 billion in the June quarter, though AI chips only account for $1 billion. AMD's CPU market share is about 34%, trailing Intel's Xeon chips.

At the event, AMD launched the fifth-generation EPYC "Turin" server CPUs, ranging from the 8-core 9015 ($527) to the 192-core 9965 ($14,831). The EPYC 9965 is said to significantly outperform Intel's flagship Xeon 8592+ server CPU.

AMD invited Kevin Salvadore, Meta's VP of Infrastructure and Engineering, to endorse their products, revealing that Meta has deployed over 1.5 million EPYC CPUs.

Disclosures

I/We may personally own shares in some of the companies mentioned above. However, those positions are not material to either the company or to my/our portfolios.