Sandia National Laboratories is pioneering advancements in high-performance computing with its latest supercomputer, Spectra, which leverages the capabilities of 128 NextSilicon Maverick-2 reconfigurable accelerators. The innovative design of Maverick-2 chips enables them to dynamically adjust their hardware configurations based on the specific demands of a workload, after initially profiling applications on a CPU. This approach eliminates the necessity for extensive software rewrites, setting it apart from traditional CPU/GPU clusters.
NextSilicon, the creators of the Maverick-2, highlight its exceptional efficiency with claims that it utilizes only half the power compared to NVIDIA’s Blackwell while doubling the speed on HPCG benchmarks and offering a tenfold increase on PageRank computations. Sandia is actively evaluating the potential of Spectra in executing molecular dynamics simulations and other Department of Energy (DOE) mission-critical codes, particularly those assignments that typically underperform on GPU-based systems, such as Monte Carlo methods.
The architecture of Maverick-2 is a fusion of Very Long Instruction Word (VLIW) capabilities, FPGA-like adaptability, and design elements reminiscent of Sony’s Cell processor. A unique compiler is employed to efficiently distribute computations across the chip’s various compute units. Despite its impressive performance metrics, the massive die size poses a manufacturing cost challenge. These chips are coupled with a RISC-V-based processor for general-purpose operations. While specific architectural details remain proprietary, the performance benefits underline a promising future for high-performance computing, forming a critical part of Sandia’s Vanguard program dedicated to exploring nascent HPC technologies for essential simulations such as nuclear stockpile stewardship.

Collections

Sandia National Laboratories has deployed Spectra, a supercomputer built around 128 NextSilicon Maverick-2 reconfigurable accelerators. The Maverick-2 chips dynamically reconfigure their hardware based on workload profiles gathered from initial CPU runs, removing the need for extensive software rewrites. NextSilicon claims the chip uses half the power of NVIDIA Blackwell while doubling HPCG benchmark performance and delivering 10x speedup on PageRank. The architecture blends VLIW, FPGA-like reconfigurability, and Cell processor-inspired design, paired with a RISC-V processor for general-purpose tasks. Sandia is evaluating Spectra for molecular dynamics simulations, Monte Carlo methods, and other DOE mission-critical workloads that traditionally underperform on GPU clusters. Spectra is part of Sandia's Vanguard program exploring emerging HPC technologies for nuclear stockpile stewardship and similar simulations.

Spectra: Redefining Supercomputing with Maverick-2 Accelerators