Intel supercomputer becomes the world’s fastest for AI

The Aurora supercomputer, developed by Intel, has become the fastest in the world in terms of artificial intelligence performance. The result was presented during the latest mixed-precision HPL-MxP benchmark.

read more

The significant milestone was reached after the system achieved an impressive 10.6 exaflops of AI performance. In this way, it did enough to surpass AMD’s Frontier, which until then occupied first place on the Top500 list of supercomputers.

Despite not yet being fully operational and facing a series of hardware problems and network instability, Aurora was not used in its entirety for the benchmark tests. Currently, the system has 21,248 CPUs and 63,744 GPUs distributed across 10,624 computing blades.

However, only 87% of the system was active during testing, reaching 1,012 exaflops and securing second place on the Top500 list. The supercomputer is currently located at Argonne National Laboratory (ANL).

Aurora has not yet fully tested Linpack

He has not yet been able to fully test Linpack for Top500 due to its operational limitations. The ANL representative indicated that there is a contractual target that exceeds Frontier’s performance, but highlighted that it is still necessary to achieve this objective so that Aurora can lead the list.

Intel highlights Aurora’s theoretical peak capacity, which reaches 2 exaflops (Rpeak), but supercomputers are evaluated by their sustained performance (Rmax). Currently, it only delivers 51% of its peak performance as sustained performance on Linpack, while Frontier delivers 70% of its total performance.

The expectation is that Aurora will improve its sustained performance over time and take the lead in the Top500 list, once it can deliver around 70% of its peak performance during sustained workloads, which would represent approximately 1.4 exaflops .

Fonte: Tomshardware