News

AMD’s AI Journey: Navigating the Data Center Landscape and Embracing Software-Hardware Synergy

  AMD’s (Nasdaq: AMD) latest fiscal 2023 third quarter report looks mixed at first glance. On the bright side, revenue for the quarter was $5.8 billion and earnings per share were $0.70, both exceeding market expectations. However, in terms of outlook for the next quarter, both indicators were slightly worse than market expectations.
  Investors naturally associate AMD with poor near-term prospects, and the stock price fell after the market closed, with the highest drop reaching 5%. However, with the holding of the earnings conference call, AMD’s stock price decline reversed. So what key information does management release in meetings?
MI300 GPU going well

  Starting in early 2023, ChatGPT set off a gold rush for generative artificial intelligence. The first beneficiary is recognized as Nvidia, because it has actually formed a market monopoly on GPU, the most critical hardware needed for the AI ​​wave. To be fair, in the first half of 2023, no company’s GPU products can compete with Nvidia’s A100/H100 GPU. Gold diggers had to pay high prices and wait in line, hoping to get the “shovel” provided by Nvidia as soon as possible.
  Against this background, the AMD MI300 series GPUs, which are expected to be launched in the fourth quarter of 2023, have high hopes from the market. Supporters of AMD believe that among the pursuers, MI300 GPU is the most promising to challenge NVIDIA GPU; while NVIDIA supporters disagree, believing that with its technology and first-mover advantage, as well as the combination of software and hardware, NVIDIA’s GPU will maintain a competitive advantage in the long term. Therefore, whether AMD’s MI300 series can be launched as scheduled and how many customers it can have after it is launched have become the focus of attention in this financial report meeting.
  AMD announced significant progress in its GPU business this quarter. On the hardware side, launch and validation of MI300A and MI300X GPUs continue as planned by the company, with performance now meeting or even exceeding the company’s original expectations. MI300A has started shipping in early October for El Capitan Exascale supercomputing. In addition, the company expects to start shipping MI300X GPUs to major cloud service customers in November (although AMD has not officially announced the list of major customers, statistics from various channels indicate that major cloud service customers include Microsoft, Meta , Oracle, Google, and Amazon; while enterprise customers include Databricks, AI21, Lamini, Moreph, and Korea Telecom).
  Based on its rapid progress in executing its AI roadmap and purchasing commitments from cloud customers, AMD expects data center GPU revenue to be approximately $400 million in the fourth quarter and to exceed $2 billion by 2024. This growth will make MI300 the fastest product in AMD’s history to reach $1 billion in sales. The announcement of this set of figures made AMD supporters excited and believed that “everything is under control.”
Continue to expand its AI software ecosystem

  In the early days of the AI ​​wave, everyone focused more on the hardware GPU field. This is of course understandable, because the hardware part is the first to generate revenue, and NVIDIA is the best example. Many people are trying to find other “shovel-selling” companies, and there are also some ambitious companies claiming that their hardware products are expected to reach Nvidia’s level of performance. But many people have overlooked one thing, that is, the popularity of Nvidia’s products is closely related to its high-level hardware, but its complete software and system ecosystem are important drivers.
  AMD obviously realizes this. In its long-term AI development strategy, software and system ecology are mentioned at a very high level. This quarter AMD announced that it continues to make progress in the AI ​​software ecosystem: the performance and functionality of ROCm software continue to be enhanced; ROCm has been fully integrated into the PyTorch and TensorFlow ecosystem; the running of the Hugging Face model on AMD GPUs has been continuously verified .
  In addition, AMD announced strategic acquisitions of Mipsology and Nod.ai during the quarter to continue to enhance its artificial intelligence software capabilities. Mipsology itself is a long-term AMD partner with proven expertise in providing artificial intelligence software and solutions for the data center, edge and embedded markets. Nod.ai focuses on providing open source artificial intelligence compiler software, which can significantly accelerate the deployment of high-performance AI models optimized for AMD processors. These software have been used by many leading cloud companies and artificial intelligence companies.
  Compared with GPU hardware sales figures, the progress in the software ecosystem may not seem so dazzling, but I am personally very concerned about it. If it wants to truly challenge Nvidia’s dominance in the field of GPU hardware, AMD’s continued progress in the software ecosystem is essential (although it may not be measured in sales for the time being).
The battle for the data center server CPU market

  Before the popularity of “artificial intelligence” and “GPU”, the battle between AMD and Intel in the field of data center server CPUs has always been the focus of market attention. AMD has continuously eroded Intel’s market share in the field of server CPUs and has been regarded by the market as maintaining performance. One of the main drivers of growth.
  When the wave of generative artificial intelligence broke out, NVIDIA CEO Jensen Huang made a famous point. He believed that data centers will rapidly shift from the current general-purpose computing based on CPUs to accelerated computing based on GPUs. The implication is that future investment in data centers will mainly fall on the GPU part, rather than the original dominant CPU. This view obviously greatly suppresses the mid- and long-term valuations of the CPU duo Intel and AMD.
  AMD CEO Su Zifeng pointed out that CPU will still have a place in the future wave of artificial intelligence. She believes that although accelerated computing will become increasingly important, GPU is not the only option. This is determined by the complexity of artificial intelligence computing tasks, and the CPU will also undertake tasks suitable for itself (in fact, Nvidia has also developed a computer code-named Grace’s CPU to compete). Therefore, AMD’s strategy is to, on the one hand, develop new GPU hardware MI300, and on the other hand, continue to firmly implement its server CPU upgrade roadmap.
  If the MI300 GPU has more of an impact on AMD’s future trend, then the sales of server CPUs will directly put pressure on AMD’s current trend. AMD gave a beautiful answer this quarter: Data center segment revenue was US$1.6 billion, flat year-on-year, but up 21% quarter-on-quarter. The main driver was the strong market demand for the third- and fourth-generation EPYC processor series, which led to quarterly Server CPU revenue hit a record high. Specifically, fourth-generation EPYC CPU revenue increased by more than 50% month-on-month, accounting for the majority of the company’s server CPU revenue and unit shipments. On the customer side, hyperscalers have expanded their deployments of EPYC CPUs to support their on-premises workloads and public instances while optimizing their infrastructure spend. Amazon, Google, Microsoft, Oracle, Tencent and others launched nearly 100 new AMD-powered cloud instances during the quarter. The company believes that AMD continued to expand its market share in the server CPU field during this quarter.
  I think AMD’s previous stock price trend was unsatisfactory. One of the important reasons is that people are worried that cloud service giants and large enterprises will give priority to GPUs (rather than CPUs) in the context of uncertain economic prospects and reduction in IT investment expenditures. But it turns out that although companies recognize the importance of GPUs (queuing up to buy them), they have not given up on CPUs. AMD’s outstanding performance in the server CPU field proves Su Zifeng’s view. In the long run, the proportion of GPUs will inevitably be an upward process, but CPUs will still have a place (rather than shrinking to dispensable).
Invest in the medium to long term

  AMD’s stock price trend is more related to the trend of changes in its overall performance. AMD’s business can be divided into four parts: data center, client (computer), game, and embedded business dominated by the original Xilinx. Although they all fall under the general framework of semiconductors, these four businesses have different cyclicalities.
  Judging from the information disclosed at the latest financial report meeting, the data center business is highly related to AI, and the upward trend in the next year is relatively clear. The driving factors mainly come from its GPU and server CPU products. The client (computer) business is closely related to the personal computer market. After years of sluggishness, the personal computer market is expected to slowly rise in the next year. The driving factors mainly come from the hardware upgrades brought by the new version of Windows operating system upgrades. , and the possible AI PC trend. However, the two businesses of gaming and embedded will enter a downward cycle. This is the unfavorable factor specifically pointed out in this financial report when looking forward to the next quarter’s performance (it is also the main reason why the company’s next quarter outlook is lower than market expectations). In particular, the embedded business has maintained strong momentum and continued to grow for several consecutive quarters, providing a tenacious driver for the company’s overall revenue when the data center and client businesses were in the doldrums; but in the end, the impact of the industry cycle cannot be completely avoided ( This time it was mainly affected by the downturn in the wireless communications industry).
  Although AMD usually only makes forecasts for the next quarter, in response to analysts’ questions, the company’s CFO analyzed the gross profit margin development trend and believed that in the second half of 2024, AMD is expected to have a good time, because it is expected to not only The data center business and client business are still on the rise, and the embedded business will also come out of the cycle trough. The upbeat statement also partially offset concerns about its short-term performance.
  Although AMD has its own AI halo and is considered to be expected to challenge Nvidia’s AI hardware dominance, I think the era of pure AI concept hype has passed. For AMD, it is unswervingly executing its AI strategy. CPUs and GPUs are advancing hand in hand, the software and system ecosystem is booming, and the revenue and profits brought by AI are becoming a reality. Therefore, from a mid- to long-term perspective, I think AMD is one of the targets worth considering.

error: Content is protected !!