Recently, at the International Supercomputing Conference (ISC 2024) held in Hamburg, Germany, the global Top500 organization officially released the 63rd edition of the Top500 list of the world's supercomputers.
Among them, Frontier, a collaboration between the Oak Ridge National Laboratory in the United States and AMD, ranked first with a peak performance of 1.206 EFlop/s. Aurora, a collaboration between the Argonne National Laboratory in the United States and Intel, broke the exascale barrier for the first time, ranking second with a peak performance of 1.012 EFlop/s.
China's supercomputers still remain out of the top ten and are no longer participating in the HPL benchmark test.
Located at Oak Ridge National Laboratory in Tennessee, USA, and operated by the Department of Energy (DOE), Frontier utilizes the HPE Cray EX235a architecture, 2GHz AMD EPYC 64C CPU, and AMD Instinct 250X GPU. It boasts a total of 8,699,904 CPU and GPU cores and utilizes the Slingshot 11 network for data transfer. Additionally, Frontier achieves an impressive rated power efficiency of up to 52.59 GFlops/W.
Compared to its initial ranking last year (when installation was not yet complete), Aurora's performance has nearly doubled. This supercomputer is housed at Argonne National Laboratory in Illinois, USA, also operated by the Department of Energy. Aurora is based on the HPE Cray EX network, integrating 21,248 Intel Xeon Max series CPUs, 63,744 Intel Max series GPUs, and 20.42PB of memory, with a peak performance claim of up to 2 EFlop/s, surpassing Frontier by a significant margin.
Installed on the Microsoft Azure cloud platform, Eagle remains the highest-ranking cloud computing system on the Top500 supercomputer list. This Microsoft NDv5 system's robust performance is attributed to the combination of Intel Xeon Platinum 8480C processors and NVIDIA H100 GPU accelerators.
Located at the RIKEN Center for Computational Science (R-CCS) in Kobe, Japan, Fugaku is based on the 2.2GHz Fujitsu A64FX 48C processor and Tofu interconnect D, boasting a total of 7,630,848 cores and maintaining its status as the highest-ranking supercomputer outside the United States.
Installed at the CSC EuroHPC center in Finland, LUMI is Europe's most powerful supercomputer. It is based on the HPE Cray EX235a system, AMD third-generation EPYC 64C 2GHz processors, AMD Instinct MI250X accelerators, and the Slingshot-11 network.
Installed at the Swiss National Supercomputing Centre (CSCS), Alps is built on the HPE Cray 254n system, featuring the 3.1GHz NVIDIA Grace 72C processor, NVIDIA GH200 superchip, and Slingshot-11 network interconnect.
Installed at the CINECA EuroHPC center in Italy, Leonardo is based on the Atos BullSequana XH2000 system, providing powerful computing capabilities with Intel Xeon Platinum CPUs and NVIDIA A100 accelerators, along with a four-rail NVIDIA HDR100 Infiniband.
Installed at the Barcelona Supercomputing Center in Spain, part of the EuroHPC initiative, MareNostrum 5 ACC is based on the BullSequana XH3000 system, featuring Intel Xeon Platinum 8460Y processors, NVIDIA H100 accelerators, and Infiniband NDR200 network connection technology.
Built by IBM at Oak Ridge National Laboratory in the USA, Summit consists of 4,356 nodes, each equipped with two 22-core 3.07GHz IBM POWER9 CPUs and six NVIDIA Tesla V100 GPUs, with each GPU integrating 80 streaming multiprocessors (SM).
Eos is an internal DGX SuperPOD used by NVIDIA, driven by Xeon Platinum 8480C 56C 3.8GHz processors and NVIDIA H100 accelerators, along with NVIDIA Infiniband NDR 400G switches.
The second-ranked Aurora supercomputer actually achieves a performance of only 1.012 ExaFlop/s, far below the claimed peak performance of up to 2 EFlop/s. Rick Stevens, Deputy Director and Distinguished Researcher at Argonne National Laboratory, explained that the 2 exaFLOPS figure is a theoretical peak based on multiplying the number of Aurora's core processors by their peak performance, but it is not the actual performance target or expectation set by the Argonne system manager responsible for maintaining and operating Aurora.
Stevens explained, "Peak performance is a theoretical number calculated based on the product of clock rate and the number of operations a single computing element can perform. In reality, typical systems can only achieve a small fraction of this peak, depending on the tasks they are performing, usually between 50%, 60%, or 70%. The 2 exaFlop/s peak claimed for Aurora is actually a combination of GPU and CPU FLOPS. This is not typically done in benchmark tests. If you look at other similar systems in the Top500 list, you will find they also have very large theoretical peaks, but the actual achieved peak ratios are similar to Aurora. Therefore, this is not a fundamental issue."
Stevens stated that Aurora should be able to achieve an improvement in LINPACK performance in benchmark tests because during recent benchmark testing, 11% of Aurora's computing power was not activated.
He predicted that higher benchmark test data will be generated once the entire system is fully installed. However, he declined to disclose the specific performance target numbers for Aurora in full operation set by Argonne National Laboratory.
According to Earl Joseph, CEO of HPC-AI industry analysis firm Hyperion Research, it is expected that over time and with further system optimization, Aurora will eventually exceed 1.5 exaFlop/s.
China no longer participates in Top500's HPL benchmark tests
After China's national supercomputers fell out of the Top500 top ten last year, they still did not make it to the latest top ten list.
In the previous list, China's Sunway TaihuLight and Tianhe-2A ranked in the top fifteen, at 11th and 14th place respectively.
The Top500 organization stated in its latest report that China has decided not to participate in Top500's HPL benchmark tests.
In the latest Top500 list, the United States remains the country with the most supercomputers. The U.S. added 7 supercomputers to its previous count, bringing the total to 168. China, on the other hand, decreased from 104 to 80 supercomputers.
The Top500 organization stated, "In fact, China did not report any new supercomputers in this new list."
Looking at the broader regional changes, North America still holds the lead, increasing from 160 systems to 171 in the previous list. In contrast, Asia saw a decline from 169 systems to 148.
Meanwhile, Europe experienced significant growth, rising from 143 to 160 systems, ranking as the second-largest region for supercomputers globally after North America.
Intel secures 63% of processor market share
The Top500 organization reported that among the top 500 supercomputers, the usage of Intel CPUs remains high at 63.00%, although lower than the 67.80% six months ago.
In comparison, 156 systems (31.20%) in the top 500 supercomputers use AMD processors, up from 28.00% six months ago.
Additionally, 194 supercomputers utilize accelerator/co-processor technology, up from 185 six months ago. Among these, 83 use NVIDIA Ampere GPUs, and 48 use NVIDIA Volta GPUs.
In terms of total computing power, the overall performance of all Top500 supercomputers increased from 7.01 ExaFlop/s to 8.21 ExaFlop/s. The performance threshold for the Top100 reached 9.44 PFlop/s.
In terms of core count, the average concurrency level of TOP500 supercomputers reached 229,426 cores per system, up from 212,027 six months ago.
Green 500 Supercomputers
There have been significant changes in this year's Green500 list, with the top three green supercomputers being newcomers.
Ranked first is JEDI–JUPITER Exascale Development Instrument, developed by EuroHPC/FZJ in Germany. JEDI ranks 190th in the TOP500 supercomputer list, demonstrating impressive performance. It boasts an energy efficiency rating of 72.73GFlop/W and an HPL score of 4.5 PFlop/s. JEDI is a supercomputer equipped with the NVIDIA Grace Hopper Superchip 72C on the BullSequana XH3000 architecture, with a total of 19,584 core processors.
Isambard-AI from the University of Bristol in the UK ranks second with an energy efficiency rating of 68.83 GFlop/W and an HPL score of 7.42 PFLop/s. Isambard-AI ranks 129th in the TOP500, with 34,272 core processors. Ranked third is the Helios system from Cyfronet in Poland. The machine achieves an energy efficiency score of 66.95 GFlop/W with an HPL score of 19.14 PFlop/s.
When discussing energy efficiency, the Top500 organization specifically mentioned the top-ranked Frontier system on the Top500 list. The Frontier system achieved an HPL score of 1.206 EFlop/s, demonstrating its exceptional computing power. Additionally, Frontier excelled in energy efficiency, scoring 56.97 GFlop/W. This places Frontier at the 11th position on the Green500 list.
