Frontier (supercomputer)
| Company type | Supercomputer |
|---|---|
| Industry | High-performance computing |
| Founded | 2022 |
| Headquarters | Oak Ridge, Tennessee , United States |
| Products | Supercomputer system (Cray EX platform) |
| Parent | Hewlett Packard Enterprise |
Hewlett Packard Enterprise Frontier, also known as OLCF-5, holds the distinction of being the first supercomputer to achieve exascale performance. Since its activation in 2022 at the Oak Ridge Leadership Computing Facility (OLCF) located in Tennessee, USA, it has been a pioneering system in high-performance computing. By November 2024, Frontier was recognized as the second fastest supercomputer worldwide. Constructed using the Cray EX platform and following the legacy of Summit (OLCF-4), it reached an Rmax of 1.102 exaFLOPS—equivalent to 1.102 quintillion floating-point calculations per second—driven by processors and accelerators from AMD.
The smaller-scale prototype, termed Frontier TDS, previously topped the Green500 rankings for energy efficiency with a performance of 62.86 gigaflops per watt until it was surpassed by the Henri supercomputer at the Flatiron Institute in November 2022.
Frontier lost its top speed status in November 2024 when it was overtaken by the El Capitan system.
Design
The Frontier supercomputer integrates 9,472 AMD Epyc 7713 "Trento" CPUs, each featuring 64 cores running at 2 GHz, totaling 606,208 cores. It also incorporates 37,888 AMD Instinct MI250X GPUs, amounting to 8,335,360 GPU cores capable of delivering double-precision operations at speeds comparable to single-precision.
The "Trento" CPU is a refined iteration of the third-generation EPYC ("Milan") processors, built on the Zen 3 architecture.
This system is housed within 74 rack cabinets, each containing 64 blades that hold two nodes apiece. Connectivity between blades is facilitated by HPE Slingshot switches with 64 ports each, providing a bandwidth capacity of 12.8 Tbps. Blades are organized into groups forming a dragonfly network topology that ensures any two nodes are separated by no more than three hops. The cabling infrastructure extends approximately 145 kilometers and utilizes both optical and copper cables optimized to reduce length wherever possible. Cooling is achieved through liquid circulation powered by four pumps rated at 350 horsepower each, moving around 6,000 gallons (22,712 liters) of water per minute without pre-chilling. This method allows for a density five times greater than traditional air-cooled systems.
Every node consists of one CPU paired with four GPUs and includes 4 TB of flash storage. Each GPU is equipped with 128 GB of RAM while CPUs feature 512 GB of local DDR4 memory. GPUs have an idle power consumption near 100 watts and reach a thermal design power (TDP) between 500 and 560 watts under full load.
A coherent interconnect architecture enables CPU code to access GPU memory directly and coherently.
Frontier's storage subsystem delivers remarkable throughput with flash read speeds of up to 75 TB/s and write speeds reaching 35 TB/s, alongside an input/output capability of 15 billion IOPS. This is supported by a site-wide Orion Lustre filesystem offering a total capacity of 700 petabytes.
The entire system operates at an estimated power draw of about 21 megawatts—sufficient to supply electricity to roughly 15,000 households—surpassing the predecessor Summit’s consumption of approximately 13 megawatts.
History
One of the primary obstacles during Frontier’s development was managing power usage efficiently. Earlier projections indicated that achieving one exaFLOP might require hundreds of thousands of GPUs and between 150 to 500 megawatts of power, making energy efficiency a critical design objective. To realize this project, Oak Ridge collaborated closely with HPE Cray and AMD, investing approximately US$600 million in construction. Installation commenced in 2021 and reached full operational status in 2022. In May that year, Frontier attained an Rmax performance exceeding one exaflop (1.1 exaflops), securing its position as the fastest supercomputer on the June 2022 TOP500 list by outperforming Fugaku.
Upon its unveiling, Frontier also led the Green500 rankings with an energy efficiency rating of 62.68 gigaflops per watt. Thomas Zacharia, then director of ORNL, emphasized that Frontier marked the beginning of a new chapter in exascale computing aimed at tackling significant scientific problems. He highlighted over ten years of joint efforts involving national laboratories, academic institutions, and industry partners under the Department of Energy’s Exascale Computing Project to ensure scalable technologies and impactful applications were successfully developed.