Supercomputing 2024: El Capitan, MI300A, and More on the Fastest Supercomputers!

Summary

At Super Computing 2024 in Atlanta, El Capitan, powered by AMD's Mi300A APU, claims the top spot in the latest Top 500 list. This unified APU integrates 24 Zen4 CPU cores and 228 CDNA3 GPU cores with 128GB of HBM3 memory, offering over 1.7 exaFLOPS of performance. Designed for high-security simulations like nuclear modeling, El Capitan sets new benchmarks in supercomputing power and efficiency. The event also highlights other powerful systems, including Intel's Aurora and Japan's Fugaku, with advancements in energy efficiency and mixed-precision performance. Exciting developments are underway in the world of supercomputing.

Takeaways

• 😀 El Capitan is the new number one supercomputer on the TOP500 list, powered by AMD's Mi 300a APU.
• 😀 The Mi 300a APU features 24 Zen4 CPU cores clocked up to 3.7 GHz and 6 CDNA 3 GPU chiplets with a peak clock of 2.1 GHz.
• 😀 El Capitan's Mi 300a APU utilizes a unified memory architecture with 128 GB of HBM3 memory and 5.3 TB/s bandwidth.
• 😀 El Capitan achieves a peak performance of 2.792 exaflops and an R-Max of 1.74 exaflops for high-performance workloads.
• 😀 The system consists of 11,136 nodes, 55,868 blades, and 87 compute cabinets, with a power draw of 29.6 MW.
• 😀 Each node in El Capitan is connected by a 200 GB Slingshot 11 network, providing up to 768 GB/s of bandwidth across the system.
• 😀 Talamine, a smaller unclassified version of El Capitan, ranks 10th on the TOP500 list with 208 petaflops of performance.
• 😀 Other notable systems include El Dorado at Sandia National Labs (ranked 20th) and RZ Adams (ranked 49th).
• 😀 The Asra 2 system in France ranks 3rd on the Green500 list for energy efficiency, scoring 69.1 gigaflops per watt.
• 😀 Aurora, an Intel-powered supercomputer, now leads in HPL Mixed Precision with 11.6 exaflops, surpassing Frontier's 11.4 exaflops.
• 😀 Fugaku continues to dominate the HPCG benchmark, holding the number one spot for four years with 16 petaflops of performance.

Q & A

What is the key feature of El Capitan's architecture that differentiates it from traditional supercomputers?

El Capitan is powered by the AMD Mi 300A APU, which is a unified APU system combining both GPU and CPU on a single chip, unlike traditional systems that separate the CPU and GPU. This design uses shared memory with unified addressing, which helps reduce memory addressing conflicts and improve overall performance.

How does the memory system in El Capitan improve its performance?

El Capitan uses 128 GB of HBM3 memory that is addressable by all the CPU and GPU chiplets in the system. This unified memory system significantly reduces memory copy operations and eliminates issues with memory address conflicts between the CPU and GPU, leading to better performance and simplified programming.

What is the significance of the SRAM used in El Capitan?

The SRAM in El Capitan, which is 256 MB in size, acts as the last-level cache for both the CPU and GPU chiplets. This shared cache allows for faster memory access and reduces latency, as it ensures both the CPU and GPU can quickly access the same data.

What is the RMax and RPeak performance of El Capitan, and how do they differ?

El Capitan has an RMax performance of 1.74 exaflops and an RPeak performance of 2.92 exaflops. RMax represents the actual sustained performance during real-world workloads, while RPeak is the theoretical maximum performance under optimal conditions.

What is the power consumption of El Capitan, and why is it significant?

El Capitan has a power consumption of 29.6 megawatts (MW). This high power draw is typical for a supercomputer of its scale, with advanced systems requiring significant energy to handle complex simulations, particularly for nuclear warhead design and high-energy physics simulations.

What are some of the primary applications El Capitan is designed to support?

El Capitan is specifically designed to run simulations for the National Nuclear Security Administration (NNSA), including nuclear warhead simulations, inertial confinement fusion, and high-energy physics simulations.

How does El Capitan compare to other supercomputers in the Top 500 list?

El Capitan holds the number one spot on the Top 500 list. It outperforms other systems in terms of computational power but is also tailored for specific applications related to national security, which is different from general-purpose supercomputers like those focused on scientific research.

What other systems on the Top 500 list are powered by the AMD Mi 300A APU?

Besides El Capitan, the **Talamine** system, located at number 10 on the Top 500 list, is another AMD Mi 300A-powered system. Talamine serves as the unclassified counterpart to El Capitan.

Which other notable supercomputers made headlines at SC24 and what were their achievements?

At SC24, **Aurora**, an Intel-powered supercomputer, was in the news for achieving the top spot in HPL mixed precision with 11.6 exaflops, ahead of **Frontier** with 11.4 exaflops. Additionally, **Fukaku** continued its dominance in the HPCG benchmark, maintaining the top spot for the fourth consecutive year with 16 petaflops.

What is the Green 500 list and how did El Capitan perform on it?

The Green 500 list ranks supercomputers based on their energy efficiency (measured in gigaflops per watt). El Capitan did not appear on this list, but another system, **Ad Astra 2** from France, ranked 3rd on the Green 500 list with 69.1 gigaflops per watt.