Due to restrictions on accessing Nvidia's most advanced GPUs, China's top 500 supercomputing systems rely exclusively on earlier-generation Nvidia accelerators, according to data from Top500.org.
According to Top500.org's Green500 ranking, which evaluates supercomputing systems based on energy efficiency (gigaflops per watt), eight of the top 10 green supercomputers utilize Nvidia accelerators, with five featuring the GH200. This contrasts sharply with the broader Top500 list, where over half of the systems lack accelerators or co-processors.
Of the top 100 green supercomputing systems, 74 systems utilize Nvidia's chips, and 13 are powered by AMD accelerators.
According to DIGITIMES Research, Nvidia is expected to account for 92.5% of the high-end AI accelerator market for cloud servers, followed by AMD's 7.3%.
China's presence on the Green500 has dwindled, with only ten supercomputing systems making the list in November 2024. This decline aligns with the broader drop in China-based systems on the Top500 list, which fell from 228 in November 2019 to 62 in 2024. Among the ten green systems, nine use Nvidia Tesla V100 accelerators, while one is powered by the Nvidia Tesla P100.
Globally, the Nvidia A100 remains the most widely used accelerator among the top supercomputing systems. However, the adoption of newer models, such as the H100 and H200, has steadily increased over the past year.
Eight of the greenest ten supercomputing systems are based in Europe, with Germany accounting for four, followed by France with two green systems.
According to The Register, Nvidia's position at the top of the Green500 supercomputing efficiency rankings could be under threat as its new Grace-Blackwell chips hit the market. The 2.7-kilowatt GB200 and the 5.4-kilowatt GB200 NVL4 represent the next step in Nvidia's product lineup, but they don't consistently deliver higher computing power per watt compared to their predecessors.
For instance, while transitioning from the A100 to the H100 between 2020 and 2022 delivered a 3.5x improvement in FP64 performance, the upcoming 1,200-watt Blackwell falls short of the 700-watt H100 in FP64 matrix math. Gains are limited to vector math, where Blackwell offers a 32% performance boost. Meanwhile, competition remains fierce—AMD's MI300A APU recently secured third place on the Green500 with the Adastra 2 system, underscoring the company's growing challenge to Nvidia's leadership.
Accelerator or co-processor utilized in top green supercomputing system | ||||
Accelerator/co-processor provider | Top 10 | Top 100 | Top 200 | Top 500 |
Nvidia | 8 | 74 | 83 | 185 |
AMD | 2 | 13 | 13 | 19 |
Intel | 0 | 2 | 3 | 5 |
Deep | 0 | 1 | 1 | 1 |
Matrix-2000 | 0 | 0 | 1 | 1 |
None | 0 | 10 | 99 | 289 |
Source: top500.org, November 2024
Number of China-based supercomputing systems utilizing accelerator or co-processor | |||||||
Accelerator/co-processor | Chip | 2022/6 | 2022/11 | 2023/6 | 2023/11 | 2024/6 | 2024/11 |
Nvidia | Nvidia Tesla V100 | 55 | 55 | 47 | 36 | 20 | 9 |
Nvidia Tesla P100 | 3 | 2 | 1 | 1 | 1 | 1 | |
Nvidia 2050 | 1 | 1 | 1 | 1 | |||
Matrix-2000 | Matrix-2000 | 1 | 1 | 1 | 1 | 1 | 1 |
Deep Computing Processor | Deep Computing Processor | 1 | 1 | 1 | 1 | 1 | 1 |
AMD | AMD Vega 20 | 1 | |||||
Intel | Intel Xeon Phi 31S1P | 1 | 1 | 1 | 1 | ||
Source: top500.org, November 2024
