On July 30th, US time, chip giant AMD was the first to release its Q2 performance report, igniting the chip stock market.
On July 31st, Nvidia (NVDA. O) surged 12.81%, Broadcom (AVGO. O) rose 11.96%, ASML. O rose 8.89%, Qualcomm (QCOM. O) rose 8.39%, TSM. N rose 7.29%, Micron Technology (MU. O) rose 7.08%, AMD (AMD) rose 8.39% O) Increased by 4.36%.
In the second quarter of 2024, AMD's total revenue was $5.835 billion, a year-on-year increase of 9% and a month on month increase of 7%; Net profit reached 265 million US dollars, a year-on-year increase of 881% and a month on month increase of 115%.
The record breaking revenue of the data center department is the backbone of AMD's strong performance growth in the second quarter. AMD's financial report shows that due to the increase in shipments of artificial intelligence GPUs, the revenue of AMD's data center division increased by 115% year-on-year to $2.8 billion.
Image source: Screenshot from AMD official website
AMD's Chairman and CEO, Zifeng Su, stated at a performance briefing that "the company's AI chip sales were 'higher than expected', with MI300 chip revenue exceeding $1 billion this quarter
The reporter from Time Weekly contacted AMD regarding issues such as market competition and core competitiveness, but as of press time, no response has been received.
MI300 breaks through the $1 billion mark
From a departmental perspective, AMD's data center division revenue increased by 115% year-on-year and 21% month on month, mainly driven by a significant increase in AMD Instinct GPU shipments and strong sales of fourth generation AMD EPYC CPUs.
The Client Business Unit benefited from the sales growth of AMD Ryzen processors, with a quarterly revenue of $1.5 billion, a year-on-year increase of 49% and a month on month increase of 9%.
However, AMD's gaming and embedded business units experienced a decline in revenue. The revenue of the gaming division for this quarter was 648 million US dollars, a year-on-year decrease of 59% and a month on month decrease of 30%, mainly due to a decrease in revenue from semi custom business. The revenue of the Embedded Business Unit for this quarter was 861 million US dollars, a year-on-year decrease of 41%, mainly due to customers continuing to adjust inventory levels, resulting in a 2% increase in revenue compared to the previous quarter.
Image source: Screenshot from AMD's financial report
The contribution of MI300 series chips is particularly significant. We delivered record breaking data center GPU revenue for the third consecutive quarter, with MI300 quarterly revenue exceeding $1 billion for the first time Su Zifeng said.
The MI300 series chips are products launched by AMD at the end of 2023, including the MI300X and MI300A chips. MI300X is a GPU accelerator, while MI300A is an APU (Accelerated Processing Unit) that integrates CPU and GPU functions.
According to a research report released by Minsheng Securities, in terms of computing power, the TF32 floating-point computing performance of MI 300X is 653.7 TFlops, the FP16 and BF16 computing performance is 1307.4 TFlops, and the FP8 and INT8 computing performance is 2614.9 TFlops, all of which are 1.3 times that of Nvidia H100. In terms of memory, the MI 300X has a memory configuration that is 2.4 times that of the Nvidia H100, with a peak storage bandwidth that is 2.4 times higher. When running Bloom, the inference speed is 1.6 times that of the H100, and when running Llama2, the inference speed is 1.4 times that of the H100. In terms of power consumption, the overall power consumption of MI300X is controlled at 750W, which is more advantageous compared to Nvidia H100. In addition, in terms of price, Su Zifeng stated that the purchase and operating costs of MI300 series chips will be lower than those of Nvidia.
During the Q2 earnings conference call, Su Zifeng mentioned that Microsoft has expanded its use of the MI300X accelerator to support GPT-4 Turbo and multiple consortium services, including Microsoft 365 Chat, Word, and Teams. Microsoft also became the first large-scale user to announce the public availability of MI300X instances.
She stated that the new Azure VM utilizes the computing performance and memory capacity of MI300X, combined with the latest ROCm software, to provide leading inference price performance when running the latest edge models (including GPT-4).
In addition to Microsoft expanding the use of AMD chips, on July 30th, Apple (AAPL. O) stated in a technical paper that two artificial intelligence models supporting its artificial intelligence system AppleIntelligence were pre trained on cloud chips designed by Google. Google's tensor processing unit was originally created for internal workloads, but is now being widely adopted. According to relevant reports, some large technology companies may be looking for and finding alternatives to NVIDIA graphics processing units in AI training.
The further maturity of AMD and Google chips has made the market realize that Nvidia's competitors are becoming stronger, and Nvidia's stock price plummeted by 7% on July 30th, evaporating nearly $200 billion in market value overnight.
Wu Quan, the director of Huaxin Jintong Semiconductor Industry Research Institute, told Time Weekly reporters that Apple's pre training on Google's cloud chips is more focused on Google's unique advantages in deep learning and machine learning, such as speed, energy efficiency, and data optimization.
Wu Quan believes that there is both competition and cooperation among technology giants, and this relationship sometimes manifests as cross competition and cooperation. For example, the collaboration between Google and Apple, AMD and Microsoft, and Nvidia and Meta. Although they intersect in fields such as artificial intelligence, GPU, and TPU, there are often differences in specific directions, which inevitably reduces direct competition and leverages their respective strengths.
The cooperation between these companies indicates that competition is misplaced in AI acceleration and large model training, and the space for cooperation is greater than competition. "Wu Quan said that the pre training process is part of their experiments, which helps explore different dimensions of AI, whether it is cloud or terminal applications. Giants are seeking differentiated development. However, this cooperation does not mean that either party will give up cooperation with other companies.
Supply chain issues remain a challenge
Su Zifeng expects that MI300 revenue will continue to increase in the third and fourth quarters of the second half of the year. AMD will continue to expand the deployment of existing customers.
Nvidia is ahead, where does AMD's advantage lie?
Su Zifeng believes that an important feature of MI300 is its leading position in the industry in terms of memory bandwidth and capacity. From a reasoning perspective, AMD's early deployment mainly focused on reasoning and performed well. Meanwhile, AMD is optimizing the ROCm software stack to make it easier for customers to train on AMD.
I expect we will continue to increase the deployment of training. Reasoning will be bigger than training, but from AMD's perspective, I believe both reasoning and training are growth opportunities, "said Su Zifeng.
In Wu Quan's opinion, AMD is more comprehensive than Nvidia. AMD offers a range of computing solutions including CPUs, GPUs, NPUs (Neural Processing Units), DPUs (Data Processing Units), and MPEs (Multiprocessing Engines), which is a significant advantage of AMD. Although comprehensiveness is an advantage, it may also bring complexity and cost challenges.
From Nvidia's perspective, its main "weakness" may lie in the CPU field. Nvidia's traditional strength lies in its GPU, but it is relatively weak in terms of CPU. However, Nvidia is also working hard to make up for this deficiency and actively expanding its CPU capabilities.
Wu Quan mentioned that AMD's MI300 series has little difference in performance compared to Nvidia's H100 series chips, and may not have an absolute advantage in comprehensiveness. However, considering the price difference, AMD chips may still be more cost-effective. It can be said that price factors have likely contributed to Microsoft and other companies making a choice between AMD and Nvidia.
Of course, Microsoft not only uses AMD chips, but also Nvidia products, which reflects the diversification of market applications. With the continuous growth of demand for AI and high-performance computing (HPC), enterprises are seeking more flexible computing power configurations. In the past, the demand for high-end computing power was widespread, but with the popularization of AI applications, the market's demand for medium or flat performance computing power is increasing. This has formed a so-called 'olive shaped' demand distribution, which means that the demand in the mid-range market is expanding, not just in the high-end market, "said Wu Quan.
However, supply chain issues remain a major challenge for AMD. Su Zifeng stated that the company is working hard to improve its supply chain situation and expects supply to continue to increase in the second half of the year. However, the overall supply chain of MI300 is tight and is expected to remain so until 2025.
AMD also announced that it will launch MI325 later this year. MI325X will utilize the same infrastructure as MI300, providing twice the memory capacity and 1.3 times the peak computing performance. In addition, AMD plans to launch the MI350 series based on the new CDNA 4 architecture in 2025, which is expected to achieve a performance improvement 35 times higher than CDNA 3 and compete with Nvidia's Blackwell solution. In 2026, AMD plans to release the MI400 series, driven by the CDNA "Next" architecture.
Su Zifeng stated that data centers will continue to be the main driving force for revenue growth in the second half of the year, and gross profit margins will continue to increase. In the long run, MI300's gross profit will exceed the company's average level. AMD expects data center GPU revenue to exceed $4.5 billion in 2024, higher than the $4 billion expected in April.