The El Capitan supercomputer will take full advantage of its nodes, each with several Instinct MI300 APU accelerators designed in sockets to be accommodated by AMD’s SP5 (LGA 6096) sockets. The system is said to be 10 times more performant than Sierra, which uses IBM’s Power9 CPUs and Nvidia’s Volta GPUs and has been running since 2018. In theoretical FP64 (double precision) computing, the system is expected to deliver an incredible 2 Exaflops of computing power and will be done at 40MW of power.
El Capitan will be designed by HPE and feature a Slingshot-11 interconnect that connects each HPE Cray XE rack. AMD is also planning to have the Instinct MI300 pave the way for future Exascale-class APUs. During AMD’s 2022 earnings day, the company revealed that the MI300 will be a multi-chip and multi-IP Instinct accelerator that not only features next-gen CDNA 3 GPU cores, but also next-gen Zen 4 CPU cores.
Designed in collaboration with AMD to deliver double-precision processing beyond 2 Exaflops, the U.S. Department of Energy, Lawrence-Livermore National Laboratory and HPE have partnered with AMD to deliver what could be the world’s fastest supercomputer, expected to be delivered in early 2023. El Capitan will be leveraged on next-generation products that incorporate improvements in custom processor design in Frontier supercomputing.
Next-generation AMD Instinct GPUs based on new compute-optimized architecture for HPC and AI workloads will use next-generation high-bandwidth memory for optimal deep learning performance. This design will excel in artificial intelligence and machine learning data analysis to create faster and more accurate models and be able to quantify the uncertainty of their predictions.
AMD will utilize the 5nm process node for its Instinct MI300 ‘CDNA 3’ GPU. The chip will be equipped with a new generation of Infinity Cache and use the fourth-generation Infinity architecture to support the CXL 3.0 ecosystem. The Instinct MI300 accelerator will feature a unified memory APU architecture and a new math format, delivering a 5x performance-per-watt improvement over CDNA 2. AMD also predicts that AI performance will increase by more than 8 times compared to the CDNA 2-based Instinct MI250X accelerator. UMAA for CDNA 3 GPUs will connect the CPU and GPU into a unified HBM memory package, eliminating redundant memory copies while providing low TCO.
AMD’s Instinct MI300 APU accelerator is expected to be available by the end of 2023, the same time frame as the aforementioned El Capitan supercomputer deployment.
