Advanced Computing in the Age of AI | Sunday, May 26, 2024

Top 500: Aurora Breaks into Exascale, but Can’t Get to the Frontier of HPC 

The 63rd installment of the TOP500 list is available today in coordination with the kickoff of ISC 2024 in Hamburg, Germany. Once again, the Frontier system at Oak Ridge National Laboratory in Tennessee, USA, retains its spot as the most powerful machine on the list with an HPL score of 1.206 exaflops.

However, this new list also shows that Frontier is no longer the only exascale machine on the list.

The Aurora machine at the Argonne Leadership Computing Facility in Illinois, USA, continues to occupy the second position on the TOP500 list. That said, the system has nearly doubled its HPL score from the previous 585.34 petaflops to 1.012 exaflops.

June 2024 Top500 results (Source: Top500.org)

The rest of the Top 10 remained largely unchanged other than the addition of the new Alps machine from the Swiss National Supercomputing Centre (CSCS) in Switzerland. This system debuted at the No. 6 spot on the list and achieved an HPL score of 270 petaflops.

This lack of change at the top of the list is indicative of a slowdown in terms of new, powerful machines being tested for the TOP500. An analysis of the list found that 20% of the Top 20 systems are more than 5 years old.

GREEN500

Unlike the TOP500, the GREEN500 saw a major shakeup for this most recent list as every machine in the TOP 3 is new.

The No. 1 spot was claimed by the JUPITER Exascale Development Instrument (JEDI), which is a new system from EuroHPC.FZJ in Germany. This machine took the No. 190 spot on the TOP500 and earned its top spot on the GREEN500 with an energy efficiency rating of 72.73 gigaflops-per-watt and an HPL score of 7.42 petaflops.

The new machine in the No. 2 slot is the Isambard-AI system out of the University of Bristol in the U.K with an energy efficiency of 68.83 gigaflops-per-watt and an HPL score of 19.14 petaflops. The Helios system from Cyfronet in Poland came in at No. 3 with an energy efficiency score of 66.95 gifgaflops-per-watt and an HPL score of 19.14 petaflops.

However, much like the previous lists, Frontier’s energy efficiency performance was so stellar here that it deserves praise. Frontier achieving the top spot on the TOP500 with an HPL score of 1.206 exaflops is impressive enough, but the system also earned an energy efficiency rating of 56.97 gigaflops-per-watt – placing Frontier at No. 11 on the GREEN500.

Despite nearly doubling its previous HPL score and officially becoming the second exascale machine on the list, Aurora simply could not match the energy efficiency of Frontier. The system took the No. 41 spot on the GREEN500 with an energy efficiency score of 26.15 gigaflops-per-watt.

Frontier-Aurora Energy Efficiency Comparison

This differential in energy efficiency between Frontier and Aurora is extremely interesting and deserves further investigation. Specifically, Frontier has about 20% more performance overall than Aurora, while Aurora is drawing 70% more power.

The core count between the two machines also makes Aurora’s current second-place spot interesting. Aurora has 9,264,128 total cores, while Frontier has 8,699,904. On top of that, both machines use the same Slingshot-11 interconnect.

Addison Snell, CEO of Intersect360 Research and Analyst Liaison at the HPC-AI Leadership Organization (HALO), discussed this on our most recent podcast. He theorized that the shared interconnect, combined with Aurora’s lower energy efficiency score compared to Frontier, could indicate an issue with Aurora’s Intel Data Center GPU Max processors.

“Standing up any machine at Exascale is a tremendous achievement, and it’s important to remember that the main point is the scientific breakthroughs the system will provide, not its Top500 ranking,” said Snell. “Nevertheless, it’s hard to get around the fact that Aurora remains short of the expectation of it becoming the world’s most powerful supercomputer.”

Sometimes referred to as the Ponte Vecchio GPU, the flagship Intel Data Center GPU Max has a Thermal Design Power (TDP) rating of 600W TDP. This is a specification that refers to the maximum amount of heat generated by the GPU that the cooling system has to dissipate. The Frontier supercomputer relies on AMD’s MI250X GPU, which clocks in at 560W TDP.

A deeper dive is required to fully understand what’s going on here, and Aurora is an exceptional machine that has officially jumped into the exascale era. However, an explanation may be that the Ponte Vecchio GPU or the GPU-CPU interconnect is demanding more power in Aurora compared to the hardware in Frontier.

Of course, Aurora is not fully finished yet, and the community expects this machine to mature as time goes on. However, as it stands now, Frontier is winning out in terms of both energy efficiency and overall power.

Hardware and the Vendor Landscape

Speaking of specific hardware, Intel Xeon processors reigned supreme in the Top 10 systems for this list. Overall, five machines use Intel Xeon Processors:

  • Aurora
  • Eagle
  • Leonardo
  • MareNostrum 5 ACC
  • EOS NVIDIA DGX SuperPod

AMD also made an appearance on the Top 10, specifically with the AMD MI250X GPU within the Frontier machine. Overall, two systems use AMD processors:

  • Frontier
  • LUMI

Additionally, one system relied on IBM processors:

  • Summit

In the past few lists, we’ve seen a shift in terms of the processors used by machines in the Top 10 positions. For the 61st list, AMD seemed to be the GPU of choice as four of the Top 10 systems relied on these types of processors (Frontier, LUMI, Perlmutter, and Selene). For that list, two machines used Intel processors (Leonardo and Tianhe-2A) and two used IBM processors (Summit and Sierra). The 62nd list saw a shift toward the current trend of using Intel Xeon processors.

In terms of interconnects, Slingshot-11 with gigabit ethernet is reigning supreme amongst the Top 10 machines. The following four machines from the Top 10 all use Slingshot-11:

  • Frontier
  • Aurora
  • LUMI
  • Alps

In terms of manufacturers, the Top 10 saw a wide spread of the usual suspects. While HPE claimed the most systems in the Top 10 (Frontier, LUMI, and Alps) EVIDEN came in second with two machines (Leonardo and MareNostrum 5.)

For the entire list, amongst the top manufacturers, the vendor share of the machines breaks down thusly:

  • Lenovo: 161
  • HPE: 111
  • EVIDEN: 49
  • DELL EMC: 33
  • NVIDIA: 23
  • Inspur: 22
  • Fujitsu: 14

For new systems added to the list, Lenovo just barely beat out the competition with 13 new systems. HPE came a close second with 12 new systems.

A Chinese Absence

Anyone who’s been following the TOP500 for the past few years will know that China has not been submitting new machines to the list – a trend that continued with the 63rd installment of the TOP500. Although China is clearly still working hard to produce competitive machines – the exascale OceanLight is an example of Chinese excellence – they simply aren’t having them tested by TOP500.

This lack of new machines has caused China’s entries on the list to slowly dry up. Although China is in second place behind the U.S. in terms of machines represented on the list by country, its losing ground. The U.S. increased its lead from 150 machines on the previous list to 161 on this one, while China once again dropped from 134 to 104.

This also led to a drop in the number of machines represented by Asia as a whole. While North America continues to have the highest number of machines represented on the list with 171 systems, Asia has dropped from 192 systems on the list to 169. Europe, on the other hand, has increased from 133 systems to 143.

EnterpriseAI