AMD is catching up in AI, has the gap with Nvidia narrowed?

Due to the prosperity of the AI ​​market, NVIDIA's market value quickly soared to US$1.04 trillion, equivalent to 5.6 times the market value of AMD. AMD will never give up and is already actively chasing this huge potential market.

In terms of chip products, AMD upgraded and iterated the Instinct product line in June and launched the Instinct MI300A, the first APU accelerator card for training large models, with 24 Zen 4 Architected CPU core, a CDNA 3 architecture GPU and 128GB HBM3 memory, the chip size reaches 146 billion transistors. At the same time, a version optimized for large language models, MI300X, was launched, which has 192GB of HBM3 memory, a memory bandwidth of 5.2TB/s, an Infinity Fabric bandwidth of 896GB/s, and up to 153 billion transistors. According to AMD official data, MI300X improves performance by 8 times and efficiency by 5 times compared to the previous generation MI250 using CDNA 2 architecture.

In terms of software ecology, AMD is also optimizing support by cooperating with more companies. In fact, compared to hardware specifications, the software ecosystem is NVIDIA's biggest competitive advantage. The CUDA architecture built over many years has long been widely accepted by the industry, with more than 4 million developers, so the vast majority of software and models can be used on the NVIDIA platform. out higher performance. AMD is currently gradually increasing the optimization of the basic framework to improve the performance of the entire platform. AMD has currently optimized its support for the PyTorch framework. According to the research results of the artificial intelligence company MosaicML, the performance of the MI250 released in 2021 has been significantly improved after optimization, and the large language model training speed has reached 80% of the performance level of the A100 graphics card.

According to predictions, the AI ​​chip market will exceed US$150 billion by 2027, and the compound annual growth rate is expected to exceed 50%. At present, Su Zifeng has elevated artificial intelligence to AMD's "highest strategic priority." If AMD can speed up the pace of catching up in AI, quickly improve hardware specifications, and further optimize the software ecosystem, will it be possible to break the current dominance of Nvidia? What about the situation?

The above is the detailed content of AMD is catching up in AI, has the gap with Nvidia narrowed?. For more information, please follow other related articles on the PHP Chinese website!