In June last year we had the opportunity to learn about the most important details of the Aurora supercomputer, a system that was developed In collaboration with HPE The goal of which was to overcome the ExaFLOP barrier, which promised to reach the maximum peak Up to 2 exaflops.
To achieve this, the Aurora has an impressive configuration, which it is equipped with 63,744 Intel Data Center Max Series GPUs, Which in turn is accompanied by neither more nor less than 21,248 processors Intel Xeon CPU Max Series. The power produced by a configuration of this type is impressive, but making everything work well and ensuring that these components can work together is more complicated.
in the end Intel and HPE have achieved this. The Aurora supercomputer is working perfectly, and has already started setting its first world records. One of them achieved this in the LINPACK mixed-precision test, where it achieved this Result of 1,012 ExaFLOPsWhich puts it in first place in the world rankings, and the best thing is that to achieve this result it did not need to use all its power, as only 9,234 knots were used, which is equivalent to 87% of the total.
The Aurora supercomputer also ranked third in the HPCG benchmark with a score of 5,612 teraflops per second (TF/s) using only 39% of the hardwarethat is, less than half of its decade, which is undoubtedly an amazing number and shows the potential that this type of equipment has.
If you are wondering why results are used at usage levels of less than 100% of the equipment, the answer is simple, because these allow you to estimate the scores and values obtained in more realistic scenarios, and are also vital Better evaluate key aspects such as communication patterns between nodes and memory access, among other things. LINKPACK offers a more comprehensive view of its power, although it doesn’t reach full saturation either.
Intel confirmed the result obtained in the mixed accuracy test under artificial intelligence It will be essential for the next generation GPU From Intel’s artificial intelligence and high-performance computing company, codenamed Falcon Shores. This will leverage the next generation Intel Xe architecture, incorporating the best of Intel Gaudi, and will allow for a unified programming interface thanks to its high degree of integration.
On the other hand, the first performance results were obtained with Intel Xeon 6 processors with P cores and 8,800 MT/s Multicast Range Memory (MCR). Show up to 2.3x performance improvement In real-world HPC applications, it is always compared to the previous generation.
“Beer enthusiast. Subtly charming alcohol junkie. Wannabe internet buff. Typical pop culture lover.”