World's fastest supercomputer crashes every few hours, performance is below design

Frontier of Oak Ridge National Laboratory (ORNL) was named the world's fastest supercomputer with a performance of up to 1,685 ExaFLOPS FP64. Frontier has a computing power of up to 1,102 ExaFlops/s, making it the first supercomputer to break the “exascale barrier”.

The Frontier supercomputer's power comes from a 64-core EPYC Trento processor, four AMD Instinct MI250X compute GPUs with 128GB HBM2E VRAM, as well as a 21MW HPE Slingshot link. The system has a total of 602,112 CPU cores and 8,138,240 GPU cores, and 4.6 PB memory of both DDR4 and HBM2e. For data transfer, Frontier is also supported by 700 petabytes of memory and Slingshot high-performance ethernet.

Theoretically, the Frontier supercomputer looks pretty good, but it has hardware problems that make it only able to deliver about 1 ExaFLOPS FP64. This disappointed many people.

According to Justin Whitt, program director at Oak Ridge Leadership Computing Facility (OLCF), Frontier is experiencing "mountains" of hardware problems with failure times in the hours, not days. Building supercomputers has always been a challenge. And it is completely understandable that a machine using a total of 60 million parts has problems.

According to rumors, the system had problems with the Slingshot connection, or AMD's Instinct MI250X compute GPU was not as reliable as expected. However, Mr. Whitt did not confirm which component the error arises from.

Whitt admitted that adjusting Frontier was "a little difficult" because of its unprecedented scale. He added that if Frontier can run more than a day without problems, that is "excellent" because the team's current goal is to operate the system by the hour.

The Frontier supercomputer has not yet been officially deployed and it is not certain whether it will be able to operate properly in 2023 as originally planned.

Google Chrome is the most vulnerable browser in 2022 with 303 security holes
Put in the crusher, how tech companies destroy millions of hard drives every year even though they are still running well
Solar storm can bring down the global Internet system
The US 'stripped' a new blow, determined to strangle Chinese semiconductors