Lawrence Livermore National Laboratory’s (LLNL) El Capitan has won the title of the world’s most powerful supercomputer. This ends Frontier’s 2.5-year reign at the top of the Top500.

With an astonishing 1.74 exaFLOPS double precision performance, El Capitan outperforms its predecessor by almost 30%. It sets a new benchmark in high-performance computing (HPC).

The technology behind El Capitan

El Capitan, built by HPE Cray, is powered by 44,544 AMD Instinct MI300A accelerated processing units (APUs). Each APU integrates 24 Zen 4 processor cores with six CDNA 3 compute dies, delivering up to 122.6 teraFLOPS of die processing. This is supported by 128GB of high-bandwidth HBM3 memory per chip, providing a blazing 5.3 terabytes per second of bandwidth.

The system operates on more than 11,136 nodes, with each node containing four MI300As connected via a 200 Gbps Slingshot-11 interconnect. Combined, El Capitan has an impressive 5.4 petabytes of memory, allowing it to handle the most complex simulations with unmatched accuracy and speed.

Bronis R. de Supinski, CTO of Livermore Computing at LLNL, explained that memory coherence between the CPU and GPU “significantly simplifies programming and optimization.”

Advancing national security and science

El Capitan’s primary mission is to safeguard the United States’ nuclear arsenal.

Corey Hinderstein of the National Nuclear Security Administration (NNSA) underlines its significance: “As NNSA’s first exascale computer, it represents a critical next step in our commitment to ensuring the safety, security, and reliability of our nation’s nuclear stockpile without the need to resume nuclear testing underground. »

In addition to its defense role, El Capitan will support groundbreaking research in various fields. From biology and weather forecasting to earthquake monitoring and artificial intelligence, the system’s capabilities will have far-reaching implications.

The supercomputer will also complement its little brother, Tuolumne, which reached 208 petaFLOPS, demonstrating the scalability of the MI300A architecture.

El Capitan: efficiency and potential

El Capitan is not only powerful but also efficient, reaching 58.89 gigaFLOPS per watt. This places it 18th on the Green500 list, a commendable position for a system of its size.

However, there is room for improvement. Currently, the supercomputer operates at 62% of its theoretical maximum performance of 2.79 exaFLOPS.

By Supinski suggested that other optimizations could be explored: “We’re not going to spend infinite amounts of cycles trying to optimize Linpack’s performance. This is not why we bought the system. I anticipate that we will likely run Linpack at scale once again, probably around the time we move the system to our classified network.

Reaching 2 exaFLOPS in real-world performance remains a potential milestone, but it would require exceeding 72% of peak efficiency or additional hardware investments.

A global perspective

Although El Capitan’s dominance is evident, the global supercomputing landscape continues to evolve. Frontier, which previously ran the Top500now faces competition from systems like the Italian HPC6 and the Swiss Alps, which have moved up the ranks.

Notably, Europe has not yet broken the exascale barrier, although the upcoming Jupiter system could change that.

Meanwhile, China remains a wild card, with reports of exascale-capable systems like Sunway Oceanlite and Tianhe-3, although no official submissions have been made to the Top500 list.

Despite this progress, El Capitan demonstrates American leadership in the field HPC. Not only does it advance technology, but it also highlights the importance of computing power to modern science and defense.