Summary of Key Points
At the Advancing AI conference, AMD unveiled a new generation of AI hardware solutions (the MI455X GPU, Venice CPU, and Helios rack system), outlining its product roadmap for the next 5-10 years. The company highlighted that its systems outperform NVIDIA's competitors in terms of performance and cost-effectiveness, with strong demand from major clients such as OpenAI and Meta. AMD predicts that the global computing market will reach $2 trillion by 2030, with AI accelerator chips playing a dominant role.
Detailed Breakdown
1. The New Generation of AI Hardware: GPU, CPU, and Rack System
AMD introduced three core products and also outlined future upgrades:
- MI455X GPU: Utilizing the most advanced 2nm/3nm chip manufacturing processes, this GPU is paired with 432GB of high-speed HBM4 memory, enabling faster and more stable AI computations. The next-generation MI500 series will be released next year, featuring even more powerful memory (HBM4e) and copper/optical interconnect technologies, resulting in an inference speed 2000 times that of the MI300X (meaning tasks that took 1 time before can now be completed in 2000 times). The MI600 series is planned for 2028.
- Venice CPU: The sixth-generation server CPU, also based on 2nm technology, generates 80% more language tokens per second when running large models compared to the previous generation. Additional products, Florence and Ravenna, are scheduled for 2028 and 2030, respectively.
- Helios Rack System: This system combines the MI455X GPU and Venice CPU into a complete “AI computing box” designed for use with large models. The Helios500 (with the MI500 GPU) will be available next year, followed by the Helios600 (with the MI600 GPU) in 2028. It can connect thousands of racks together to support extremely large AI models.
2. Competing with NVIDIA: AMD’s New Systems Are Faster and More Cost-Effective
AMD directly compared its new systems with NVIDIA’s Vera Rubin NVL72 rack system, highlighting three advantages:
- Improved Performance: The AI computing power (in FP4 precision) is over 15% higher, and the memory capacity and speed are more than 50% greater. When running the Kimi2 large model, the throughput increases by 15% in scenarios involving 32,000 input words and 8,000 output words.
- Better Cost-Effectiveness: For every dollar invested, AMD’s systems generate 30% more tokens, making them more cost-efficient for businesses.
3. Major Clients Show Strong Interest
AMD’s new systems have been embraced by top clients:
- OpenAI: Will start using Helios by the end of this year and plans to deploy it on a large scale in 2027, with the intention of transitioning to the next-generation MI500 GPU.
- Anthropic: Its Claude large model adapted to AMD’s rack system in just one weekend, demonstrating exceptional performance.
- Other Clients: Meta, Microsoft, and Oracle are also among AMD’s partners. Helios is already in mass production, with shipments expected by the end of the third quarter, and demand from these clients is very high.
4. Long-Term Vision: Aiming for the $2 Trillion Computing Market by 2030
AMD has set ambitious market goals:
- Market Size: The global computing market is expected to reach $2 trillion in 2030, with AI accelerator chips accounting for $1.4 trillion (70%) and CPUs for $220 billion (11%); AI will be the dominant force.
- AMD’s Own Goals: Aim to capture 46% of the server CPU market share by 2026, competing with Intel and NVIDIA.
- Collaborative Approaches: Working with AI chip company Cerebras, AMD is developing inference solutions that combine Helios with wafer-level chips to tackle more complex AI tasks.
5. Why Are Rack-Level Systems in High Demand for AI?
The popularity of rack-level systems stems from the explosive growth in AI demand:
- Dramatic Increase in Demand: The monthly consumption of tokens processed by AI has increased by 158 times in two years, and computing demands have grown fivefold since 2020. Traditional single-chip solutions can no longer meet the needs of large models.
- Integration Benefits: Integrating GPU, CPU, and memory functions into a single rack enhances efficiency, supports longer contexts (e.g., allowing AI to retain more previous conversations), and enables easy expansion to thousands of racks to accommodate extremely large models.
In One Sentence
AMD’s new hardware and systems not only challenge NVIDIA in terms of performance and cost-effectiveness but also aim to become a key player in the future AI computing market through partnerships with major clients and long-term strategic planning, positioning itself as a core component of AI infrastructure.