Analysis of AMD’s Technical Advantages in the Global Top500 Supercomputers Ranking

Introduction

In November 2024, the supercomputing landscape experienced significant changes, with AMD making substantial progress in the Top500 rankings. This development is mainly reflected in the deployment of the El Capitan supercomputer.

Figure 1: A detailed view of modern supercomputer core computing hardware.

El Capitan System Architecture

The El Capitan system is equipped with 43,808 AMD "Antares-A" Instinct MI300A devices, showcasing remarkable computational power. In the High Performance Linpack (HPL) test, the system achieved a theoretical peak performance of 2,746.4 petaflops, with a sustained performance of 1,742 petaflops, resulting in an efficiency of 63.4%.

Figure 2: Distribution of computing architectures in the Top500 list, with AMD Instinct (green), NVIDIA Grace (brown), and other architectures represented by different colors.

Technical Specifications

The MI300A's technical specifications feature three chiplets integrated with 24 "Genoa" Epyc cores, and six GPU stream processing chiplets operating at 1.8 GHz. The system architecture uses HPE's "Rosetta" Slingshot 11 Ethernet interconnect and integrates more than 1.05 million Genoa cores and approximately 10 million GPU stream processors.

Figure 3: Comprehensive system specifications, performance metrics, and core count of various supercomputers.

Market Performance and Competition

The November 2024 rankings revealed that AMD’s market penetration has significantly increased, with AMD GPUs contributing 72.1% of the newly added performance. This includes the El Capitan system and three smaller systems based on the MI300A hybrid computing engine, collectively delivering 3,134.6 petaflops (FP64 performance), accounting for 60.1% of newly added computing power.

Figure 4: Statistics of supercomputers categorized by accelerator type, including system count, peak performance, and core distribution.

High Performance Trends

Figure 5: Historical performance trends of supercomputers from 1990 to 2025, displaying total computing power, the first-ranked system, and the 500th-ranked system on a logarithmic scale.

Competitive Landscape Analysis

The competitive landscape has undergone noticeable shifts. NVIDIA’s share in newly deployed systems has declined. Grace-Hopper systems accounted for only 3.8% of the newly added computing power, while systems combining Intel Xeon CPUs with NVIDIA GPUs contributed 18.6%.

Industry Impact

From a broader perspective, the HPC industry is undergoing major transformations. The total computational power of all 500 systems has reached 1.172 exaflops, a significant increase from 0.821 exaflops in June 2024. However, maintaining historical growth rates at the lower end of the Top500 rankings remains challenging.

These developments indicate that the supercomputing sector is in a transition phase, with AMD’s technological advancements reshaping the competitive landscape. The success of hybrid CPU-GPU architectures in achieving exascale performance provides new insights into next-generation supercomputer system designs.

Currently, the minimum threshold for entering the Top500 list is 2.31 petaflops (HPL benchmark), while reaching the Top100 requires 12.8 petaflops. This progress highlights improvements in high-end system capabilities and the ongoing challenges in scaling HPC performance.

Conclusion

The evolution of supercomputing architectures and performance metrics demonstrates the industry’s continuous pursuit of greater computational power. At the same time, it raises questions about the sustainability of current growth trends and the role of cloud computing in future HPC deployments.

References

[1] T. P. Morgan, "AMD Is The Undisputed Datacenter GPU Performance Champ – For Now," The Next Platform, Dec. 6, 2023. [Online]. Available: https://www.nextplatform.com/2023/12/06/amd-is-the-undisputed-datacenter-gpu-performance-champ-for-now/ [Accessed: Dec. 5, 2024]