On Thursday, October 10th, $Advanced Micro Devices (AMD.US)$ Chairman and CEO Lisa Su introduced the latest flagship AI chip - AMD Instinct MI325X GPU (referred to as AMD MI325X), intending to directly compare with NVIDIA's popular product H200. $NVIDIA (NVDA.US)$ , and also thoughtfully compared AMD MI325X with NVIDIA's hot selling product H200.

Su said at the conference that AMD MI325X adopts 256GB HBM3E supporting 6.0TB/s, providing 1.8 times more capacity and 1.3 times more bandwidth than NVIDIA's H200. Compared to H200, AMD MI325X's peak theoretical FP16 and FP8 computing performance has increased by 1.3 times.

In terms of AI training, the performance of the AMD Instinct MI325X platform is no less than NVIDIA's H200 HGX. As shown in the graph, whether in a single GPU or in an 8 GPU Meta Llama-2 training scenario, the performance of the AMD Instinct MI325X platform is on par with the latter, giving the company more leverage in this market.

AMD宣称，AMD MI325X有望在 2024 年第四季度投入生产，预计从 2025 年第一季度开始将在包括戴尔、Eviden、技嘉、惠普、联想等在内的众多平台提供商的系统中广泛使用。

苏姿丰还在会上公布了最新的AI芯片路线图，采用该公司CDNA 4架构的MI350系列明年上市，MI400系列将采用更先进的CDNA架构。

AMD通过发布MI300X和MI300A等性价比高的AI芯片，受到一些行业巨头的青睐。苏姿丰称，$Microsoft (MSFT.US)$、OpenAI、 $Meta Platforms (META.US)$Please use your Futubull account to access the feature.$Coherent (COHR.US)$ Multiple manufacturers' generative ai platforms have used the MI300 series driver, indicating that AMD may have the opportunity to gain a foothold in this competitive market.

However, the AMD stock feedback on the day of the conference was not very positive, closing at $164.18 on Thursday, down 4%, but rising 2.26% the next day to close at $167.89.

Surpass Nvidia? Or earn money first

Analysts comment that although AMD has released new chips, AMD is far behind the industry leader Nvidia, with the company's recent financial outlook showing little change, and investors are still waiting for returns from the AI business.

KeyBanc's analyst John Vinh believes that despite AMD lagging behind Nvidia, AMD's development trajectory is clearly optimistic, with shipments of 0.5 million Instinct MI300X AI accelerators expected this year, further gaining a foothold in the AI market.

Lynx Equity Strategies analyst is skeptical about whether AMD can narrow the gap with Nvidia, unless AMD can demonstrate a clear market share growth driver, otherwise the stock may retest its annual low.

Wall Street has been waiting for signs of AMD catching up with Nvidia, this progress may have to wait until AMD's next financial report, which is expected around the end of this month, after AMD releases its third quarter financial report for this year.

Specifically, looking at the past three quarters, AMD datacenter FY2023Q4, FY2024Q1, FY2024Q2 revenues were $2.282 billion, $2.337 billion, $2.834 billion; while Nvidia datacenter FY24Q4, FY25Q1, FY25Q2 revenues were $18.4 billion, $22.675 billion, $26.272 billion respectively.

In terms of quarter-on-quarter revenue growth, Nvidia's datacenter quarterly revenue growth has declined, with AMD catching up with a 20% quarter-on-quarter growth.

According to the market consensus, it is expected that AMD Datacenter FY2024Q3 revenue will be $3.455 billion, while Nvidia Datacenter FY2025Q3 revenue is expected to be $29.005 billion.

Nvidia leads the market in datacenter revenue, thanks to its "unrivaled" GPU market share. According to TrendForce's AI Server industry report released in July this year, Nvidia holds the highest market share in the AI server market equipped with GPUs - close to 90% - while AMD's market share is only 8%.

However, including all AI chips (GPU, ASIC, FPGA) used in AI servers, Nvidia's market share this year is about 64%; AMD's AI chip market share has significantly increased from 5.70% in 2022 to 8.1% in 2024.

Despite the increase in market share, AMD still faces Nvidia's biggest challenge - the programming language CUDA. The CUDA platform greatly reduces the difficulty of using Nvidia GPUs for high-performance computing, providing developers with a powerful toolkit that makes it easier for them to utilize GPUs for various compute-intensive tasks. This has become the standard for AI developers, effectively locking developers into Nvidia's ecosystem.

In response, AMD has been improving its competitive software, ROCm, so that AI developers can easily switch more AI models to AMD chips.

Kicking Intel, AMD's ambitions are beyond just GPUs.

As the AI landscape unfolds, AMD's ambitions in AI go beyond GPUs to surpass Intel in the datacenter CPU field.

Su Zifeng stated that AMD EPYC CPU's market share in the datacenter CPU market has climbed to 34%, which was only 2% in 2018, but still lower than Intel. To change this situation, AMD plans to challenge Intel's market position with the new CPU series EPYC 5th Gen.

It is a good time for AMD to catch up with Intel, which is currently in a weak position. $Intel (INTC.US)$ Not only facing continuous decline in business revenue, Intel is also experiencing its worst financial condition in 50 years. The lack of AI business opportunities is the main issue, rather than the poor execution of foundry and chip department goals.

Su Zifeng emphasized that today's artificial intelligence is actually related to CPU capabilities. This can be seen in data analysis and many other similar applications. The newly released CPU is particularly suitable for processing data into AI workloads, as almost all GPUs require a CPU on the same system to start computations.

Clearly, AMD has emerged victorious in the competition with Intel, but no more.

In August of this year, AMD announced its plan to invest $4.9 billion to acquire server manufacturer ZT Systems. ZT Systems' rich experience in designing and optimizing cloud computing solutions will significantly accelerate the large-scale deployment of AMD-driven AI infrastructure for cloud and enterprise customers.

Su Zifeng mentioned the acquisition of ZT Systems, expressing the hope to lay out the artificial intelligence software stack. With this acquisition, AMD will integrate various elements to provide a true artificial intelligence solution roadmap.

Of course, this is based on AMD's optimistic outlook on the datacenter market. "The market for datacenter AI accelerators is projected to reach $500 billion in 2028, up from $45 billion in 2023," said Su Zifeng at an AI conference on Thursday.

During the earnings call in July this year, Su Zifeng stated that the AMD Instinct MI300X GPU generated over $1 billion in revenue for datacenter in the second quarter. She also mentioned that these accelerators are expected to bring in over $4.5 billion in revenue for AMD in the 2024 fiscal year, surpassing the $4 billion target set in April.

Despite tough competition, Su Zifeng previously mentioned in an interview that the AI chip market is big enough to accommodate multiple companies, and AMD does not have to defeat Nvidia to be successful.

Editor/Rocky