AMD's MI300 can co-exist with ASICs and GPU solutions in the AI chip market, CEO Lisa Su has said recently, stressing that its latest AI chip is very powerful and gaining market traction. She also believes that GPUs remain the preferred processing solution for AI workloads, while ASIC chips can be used for specific tasks where efficiency is critical.
Su made the remarks at the firm's October 31 earnings call conference for the third quarter of 2023 when asked by institutional investors about whether MI300, whose main workload is AI inference, would be replaced by ASIC chips developed by cloud customers over time.
At the conference, AMD's MI300 was mentioned 20 times, but no investors asked about AMD's AI chip orders from China. The management team also did not proactively mention the impact of the newly updated ban on AI chip exports to China. This indicates that the ban does not appear to have a significant impact on the MI300 series of chips, given that publicly traded US companies are required to proactively report any major events that could impact their stock prices.
Compared to Intel's aggressive expansion of its AI chip Gaudi 2 solution in the Chinese market, AMD's current AI chip customers are still mainly in the North American market. Earlier, foreign media reported that the MI300 series had won orders from Oracle, while also gaining favor of Microsoft.
Su noted that AMD's data center GPU revenue for 2024 could reach US$2 billion, driven by AI demand. She emphasized that AMD has been planning supply chain capacity since 2022, and therefore can fulfill robust orders in hand.
MI300 gaining ground in the AI chip market
Market analysts said that AMD's MI300 series of AI chips is gaining traction in the market, but the amount is small compared to the tens of billions of dollars in AI chip orders that NVIDIA has lost in China following the escalated chip ban.
This also suggests that AMD's plan to take advantage of Nvidia capacity constraints and poach customers will take time to materialize, the analysts said.
At the latest earnings call, multiple investors also asked about the progress of AMD's MI300 series, focusing on the following aspects: customer acquisition, capacity supply, and whether it will be marginalized by cloud customers' self-developed chips.
In response, Su said that AMD has a wide range of customers in the AI field, from hyperscale cloud providers to OEMs, enterprise customers, and some AI startups. From a workload perspective, MI300 can handle both AI training and inference workloads, especially for large language model (LLM) inference tasks. This is a key advantage for AMD, as LLMs are becoming increasingly important in a variety of applications.
Meanwhile, market observers noted that in light of the US government's updated chip trade ban on China, Nvidia, Intel, and AMD have taken different approaches to address the new regulations and their potential impact.
Nvidia immediately complied with the ban, suspending sales of its high-end AI chips to the Chinese market. This resulted in a loss of over US$5 billion in chip orders from Chinese customers like ByteDance, Tencent, and Alibaba.
Intel, on the other hand, has maintained a more optimistic stance, emphasizing its close communication with the US Department of Commerce and its belief that it can continue selling chips to China. Intel is hoping to expand its sales of its AI chip Gaudi 2 solution in the Chinese market, despite the ban.
AMD has taken a more cautious approach, avoiding any public statements about the impact of the ban on its AI chip business.
Article translated by Willis Ke
