In a nutshell: Aurora, the supercomputer created by way of a collaboration between Intel, HPE, and the US Division of Vitality’s Argonne Nationwide Laboratory, has damaged the Exascale barrier, changing into the second machine after Frontier to hit that coveted mark. It’s now additionally the quickest AI supercomputer on the planet, though it nonetheless did not dislodge Frontier from its perch because the quickest machine on this planet.

Aurora debuted in second place on the Top500 listing final November, however again then, it solely managed to get Linpack working on about half the system. Even with solely partial deployment, it supplied 585 petaFLOPS of double precision efficiency. The newest rating offers us a greater thought of the supercomputer’s capabilities. Nevertheless, it’s price noting that the machine continues to be not one hundred pc practical, that means the benchmarks don’t replicate the system’s full potential.

Aurora has 21,248 Intel Xeon CPU Max processors and 63,744 Intel Information Middle GPU Max accelerators. Within the newest exams, it spat out 1.012 exaflops of efficiency with solely 9,234 nodes energetic out of the entire 10,624. Whereas it’s awe-inspiring when taken in isolation, it is nonetheless disappointing for the much-hyped and much-delayed system that the majority anticipated to finish the AMD-powered Frontier’s reign because the quickest supercomputer on this planet.

Aurora carried out a lot better within the HPL-MxP mixed-precision benchmark, registering 10.6 exaflops of AI efficiency with solely 89 p.c of the entire system energetic. Not like Linpack, which makes use of the FP64 precision, HPL-MxP makes use of FP32 and FP16, which higher symbolize a machine’s AI capabilities. Whereas FP64 is generally used for conventional scientific computing, AI workloads and different real-world functions primarily use decrease precision, that means Linpack is not the one indicator of the true capabilities of a supercomputer.

Argonne’s lack of ability to run benchmarks on the total Aurora {hardware} is because of quite a few system failures, together with cooling points, operational errors, and community instability. Nevertheless, even when engineers repair these issues and completely tune the system, it’s nonetheless unlikely to outstrip Frontier.

The AMD-powered system is at present rated for 1.206 exaflops of efficiency, round 19 p.c sooner than Aurora, with 87 p.c of the latter’s {hardware} up and working. Even with linear scaling, the remaining 13 p.c of the Aurora nodes will not add the 200-odd petaflops wanted to topple the present chief. Argonne representatives nonetheless insist that Aurora “ought to” beat Frontier when totally deployed, however whether or not that occurs stays to be seen.