At AMD's "Advancing AI" event in 2025, CEO Lisa Su made a remarkable prediction: the global AI data center accelerator market will skyrocket from $71 million to $500 billion within three years. This forecast not only reflects the explosive growth trend of AI technology, but also reveals AMD's strategic ambition and confidence in the new round of computing revolution. China Exportsemi will conduct an in-depth analysis of how AMD grasps the dividends of this era from the aspects of computing power demand, technology layout, market competition and challenges.
Ⅰ. AI has driven the demand for computing power to skyrocket
1. The model iteration detonates the computing power demand
Over the past five years, the size of AI models has grown exponentially, from 340 million parameters in Google BERT in 2018 to trillions of parameters in OpenAI's GPT-4 in 2023. This growth means that the computing power required to train a large model far exceeds that of traditional server CPUs, and GPUs and AI accelerators have become the main force. For example, GPT-3 requires 3640 PF-days of floating-point operations for a single training, and if you use a high-end GPU cluster such as MI300X, you can significantly reduce training time and energy consumption.
2. Multi-scenario penetration accelerates the popularization of AI
The application scenarios of AI are moving from the "center" to the "edge".。 In the cloud, large model inference services such as ChatGPT and Copilot continue to explode; On the terminal and edge side, the demand for real-time inference is becoming more and more intense for smart cars, industrial robots, smart factories, and medical image analysis. According to IDC, the global edge computing market will reach $176 billion in 2023 and is expected to exceed $300 billion by 2026, with a compound annual growth rate of more than 15%.
3. Data bursts drive intelligent processing upgrades
Global data volumes continue to explode. IDC predicts that by 2025, the total amount of data generated worldwide will reach 175 zettabytes. At the same time, the number of IoT devices will double to 30 billion by 2028. With such a large flow of data, enterprises urgently need to upgrade their data centers to achieve more efficient training and inference tasks. As the core computing power unit, the demand for AI accelerators has risen.
Figure: AMD CEO Lisa Su is strongly bullish on the AI data center market
Ⅱ. AMD's technology and marketing strategy
1. Rapid upgrade of hardware product line
AMD's MI350 series accelerators are based on the CDNA 3 architecture and TSMC's 3nm process, focusing on strengthening energy efficiency and throughput in large model inference scenarios. The flagship model MI355X consumes 1400W of power and is equipped with the latest HBM3E high-bandwidth memory, with a total bandwidth of more than 5TB/s, capable of carrying real-time inference of tens of billions of parameter-level models. Up to 35x faster inference performance than its predecessor.
What's more noteworthy is that AMD has made it clear that its MI300 series GPUs have been deployed in multiple supercomputing centers, and will form a "two-wheel drive" with the MI350 series in the future, connecting the two major scenarios of training and inference.
2. The software ecosystem continues to consolidate
No matter how strong the hardware is, it also needs the collaboration of the software ecosystem. AMD has been promoting the evolution of the ROCm (Radeon Open Compute) platform in recent years, and has now updated to ROCm 7. The platform is compatible with mainstream AI frameworks (such as PyTorch and TensorFlow), and introduces inference engines optimized for large models, such as vLLM and llm-d. These tools can significantly reduce memory usage and increase throughput, helping developers deploy models more efficiently.
In addition, AMD also promotes the construction of an open-source ecosystem, and cooperates with leading developers such as Hugging Face, OpenAI, and Meta to jointly optimize model adaptation and form a virtuous ecological cycle.
3. Build a full-stack solution
Unlike a point product strategy, AMD is accelerating the construction of an integrated approach from chip design to system deployment. This includes not only GPU and CPU collaboration (e.g., EPYC and MI series complement each other), but also underlying compilers, drivers, schedulers, and containerized management tools. AMD has partnered with a number of CSPs (Cloud Service Providers) to build a full-stack AI service platform to provide enterprise customers with out-of-the-box AI capabilities and lower the threshold for application.
Ⅲ. Competitive landscape and market opportunities
1. Differentiated competition against NVIDIA
In the field of AI accelerators, Nvidia still occupies the absolute leading position, but AMD's market share is growing rapidly. According to Jon Peddie Research, AMD's data center GPU market share reached 21.6% in the first quarter of 2024, compared to a single-digit market in the same period last year. AMD leverages the synergies between EPYC processors and the MI series to differentiate itself in terms of performance, cost, and energy consumption.
In addition, compared with NVIDIA's closed ecosystem with CUDA as the core, AMD tends to build an open-source and customizable technology stack to attract a group of enterprise users who want to avoid "lock-in".
2. Expand cooperation with cloud vendors and break down channel barriers
AMD is working with major global cloud service providers, including Microsoft Azure, Google Cloud, Oracle Cloud, and Tencent Cloud, to provide a cloud deployment environment for the MI series accelerator cards. Through the in-depth cooperation with CSPs, AMD can not only obtain customer feedback and optimize products more quickly, but also accelerate the market share of cloud inference and training.
Ⅳ. The challenges remain
1. There is a lot of pressure on technology iteration
The speed at which AI technology is evolving is astonishing. From Transformer to MoE (Hybrid Expert Model) to RAG (Retrieval Enhanced Generation) and Agent architecture, the hardware requirements for algorithms are also increasing. AMD must continue to invest in R&D in memory bandwidth, interconnect architecture, power consumption control, and more to stay ahead of the curve.
2. The influx of new players has intensified the involution
In addition to NVIDIA, startups such as Graphcore, Cerebras, and Groq have released AI-specific chips, emphasizing higher energy efficiency and scenario customization. Some products have capabilities that surpass traditional GPU architectures in specific tasks (such as token-level inference and Sparse Computation). AMD faces both technical and market challenges from emerging players.
3. Uncertainty in the supply chain and capacity
High-end accelerators are highly dependent on advanced manufacturing processes and HBM's supply chain. If foundries such as TSMC and Samsung are tight, or HBM's supply chain such as SK hynix is limited, it will directly affect AMD's product delivery rhythm. In addition, geopolitics, export controls and other variables may also become obstacles to AMD's global layout.
Ⅴ. Conclusion: The race to $500 billion
AMD's optimistic outlook for the AI data center market is considered to be aggressive, but it is not unfounded based on the speed of its product iteration, ecosystem layout, and breadth of customer cooperation. AI has entered thousands of industries from the laboratory, and the demand for computing power has entered a stage of exponential growth, and whoever can provide a more cost-effective acceleration solution will be able to take the initiative in this "second Moore's Law race".
In the next three years, it remains to be seen whether the MI350 series and subsequent products can truly shake NVIDIA's dominance, but it is certain that AMD is no longer a "CPU-only company", but a key player actively fighting for a new round of voice in the AI era.
