The U.S. Energy Department’s Aurora supercomputer has achieved the highest score yet in a closely-watched artificial intelligence performance test.

The speed evaluation was carried out by the TOP500 supercomputing organization, which published its findings today. The group also tested Aurora’s performance across other workloads besides AI applications. It determined that the system is the second-fastest overall behind Frontier, another Energy Department supercomputer. 

Aurora comprises 166 refrigerator-sized cabinets that each hold 64 compute modules (pictured). Each compute module weighs 70 pounds and includes eight Intel Corp. chips: six graphics processing units backed by a pair of central processing units. The assembly of Aurora’s hardware components was completed last year, but the supercomputer is not yet fully operational.

TOP500 evaluated the system’s AI performance using a benchmark known as HPL-MxP. It measures how supercomputers process floating point numbers, units of data that lend themselves well to representing long fractions and are used extensively by AI models. Aurora achieved 10.6 exaflops of performance, a new record, even though only 87% of its compute nodes were active during the test.

The Energy Department’s other exascale system, Frontier, came in second with 10.2 exaflops. But it retained its top spot as the fastest supercomputer overall.

TOP500’s flagship performance test compares supercomputers using a benchmark called HLP. It measures the evaluated systems’ speed by analyzing how fast they can solve systems of linear equations. Frontier achieved an HLP score of 1.206 exaflops while Aurora managed 1.012 exaflops.

Both supercomputers were built by Hewlett Packard Enterprise Co. for the Energy Department. But whereas Aurora uses Intel silicon, Frontier is powered by processors from Advanced Micro Devices Inc. The latter system comprises 74 cabinets that hold 9,400 servers, each of which features one CPU and four AI accelerators.

Aurora could potentially take the first spot from Frontier on Top500’s flagship HLP ranking once it becomes fully operational. An Intel spokesperson told Tom’s Hardware that “there’s a contractual target number that is faster than Frontier. So, if we’re successful in reaching that number, we’ll be faster than Frontier.”

The next three spots on today’s TOP500 ranking are held by the same systems that rounded out the top five list last November, when the last round of benchmark tests took place. 

The world’s third-fastest supercomputer behind Frontier and Aurora is Eagle, an AI-optimized system Microsoft Corp. installed in its Azure cloud platform last year. It’s powered by H100 graphics cards from Nvidia Corp. and Intel CPUs. The fourth spot went on to Arm-based Fugaku supercomputer, which is located at the Riken Center for Computational Science in Japan, while Finland’s LUMI system rounded out the top five.

Photo: Intel