Reconfigurable Supercomputer Offers Potential Leap in Computational Efficiency
Sandia National Laboratories has initiated testing of a novel type of supercomputer that distinguishes itself from conventional designs. Unlike traditional supercomputers built with interconnected clusters of CPUs and GPUs, this new machine incorporates reconfigurable accelerators. These accelerators are optimized for specific computational tasks, offering a significant advantage in efficiency. The architecture, developed by startup NextSilicon, bears resemblance to field-programmable gate arrays (FPGAs). A key feature of this approach is its ability to optimize hardware for software, eliminating the need for extensive software rewrites.
According to James Laros, program leader and senior scientist at Sandia, the experimental supercomputer, named Spectra, incorporates 128 NextSilicon Maverick-2 accelerators. NextSilicon, headquartered in Tel Aviv and Minneapolis, claims its accelerators can consume up to half the power of NVIDIA’s Blackwell processors while delivering a fourfold speed advantage, although these figures can vary depending on the specific application. Elaad Raz, CEO of NextSilicon, explained the rationale behind the design, stating that traditional architectures often involve predicting the next instruction and caching data. He questioned the overhead associated with this approach, aiming to create a system where hardware and software collaborate to optimize performance.
The Spectra architecture first analyzes an application on a CPU to identify frequently executed operations. It then reconfigures the chip to schedule work in a manner that optimizes data flow, creating a pipeline to reduce data fetching overhead. A significant benefit is that users do not need to modify their software to take advantage of the system’s capabilities; the hardware adapts to the software. Laros noted that many applications Sandia runs are limited by memory bandwidth, and Spectra’s architecture has the potential to overcome this limitation by reducing the need to access the main memory. Raz reports that the Maverick-2 can perform HPCG, a supercomputing benchmark, twice as fast and PageRank, another benchmark, ten times as fast.
Sandia scientists are currently evaluating Spectra’s performance in molecular dynamics simulations, which are crucial in physics and materials science, and other core codes used by the U.S. Department of Energy. Laros highlighted that the system’s potential lies in improving performance for applications that do not run well on GPUs or in achieving the same performance with greater energy efficiency. Sandia’s mission involves maintaining the nation’s nuclear arsenal, making it imperative to have diverse and robust computational capabilities. The lab’s Vanguard program facilitates partnerships with startups like NextSilicon to test and develop advanced computing technologies. This approach ensures that the U.S. government has alternative options should a technology provider face challenges.
While GPUs offer speed advantages, porting existing code to these systems can be a time-consuming process. Furthermore, some scientific simulations, such as Monte Carlo methods used for risk assessment, do not perform optimally on GPUs. Laros observed a current trend in computing startups focusing on artificial intelligence, but NextSilicon’s technology holds promise for both AI and traditional high-performance scientific computing due to its power efficiency. Power availability is a major constraint for large-scale AI data centers, and NextSilicon’s accelerators aim to address this by enabling more efficient performance per unit of electricity consumed.
The Vanguard program’s experimental nature acknowledges the inherent risk in exploring cutting-edge technologies. Laros emphasized that the goal is to pioneer advanced technology discovery, with the expectation that other labs and commercial industries will follow suit. The successful testing of Spectra could pave the way for greater innovation in high-performance computing and address the growing demand for energy-efficient solutions in data centers and scientific research.
