Maintaining technical superiority requires an awareness of the newest hardware capabilities as artificial intelligence alters industry and national security.
For almost ten years, AI accelerators—specialized systems intended to accelerate tasks like neural networks, deep learning, and machine learning—have been a significant field of development. The Lincoln AI Computing Survey (LAICS, pronounced “lace”) has been carried out since 2018 by a team from the Lincoln Laboratory Supercomputing Center (LLSC). After six articles, LAICS is still summarizing and comparing the peak power and performance of existing commercial AI accelerators. The number of research AI accelerators reported in research publications and commercial accelerators launched increased dramatically about eight years ago, and government supporters of the laboratory’s work began asking us about them. According to Albert Reuther, a staff member of the LLSC, which manages and improves the high-performance computer systems utilized by hundreds of laboratory research personnel, “that was motivation enough to start the survey.”
While AI accelerators are commonly employed for tasks like machine learning, they can also facilitate other parallel applications, such as modeling molecular functions and accelerating fluid dynamics simulations, which are computationally costly.
Central processing units (CPUs), graphics processing units (GPUs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), and dataflow accelerators are some examples of AI accelerator technology. The capabilities of each kind of accelerator vary slightly. While ASICs are limited to extremely specific jobs, CPUs can be utilized for general-purpose processing. GPUs, FPGAs, and dataflow accelerators are more adaptable and can be set up for a range of tasks. Depending on their design, the various types of accelerators have different levels of efficiency and performance. In order to identify the best accelerators for certain purposes, LAICS compares and surveys the technologies currently available on the market.
Michael Jones, Peter Michaleas, Jeremy Kepner, and Vijay Gadepally are LLSC members who are part of Reuther’s LAICS team. In order to understand how accelerators help research and development for their goals, the team also works with researchers from several divisions within Lincoln Laboratory, such as the Advanced Technology Division and the Intelligence, Surveillance, and Reconnaissance and Tactical Systems Division.
The most recent study in their series examined over 120 accelerators, whereas the first one examined 57 accelerators. The team sorts accelerators based on whether they are on a chip, card, or system after comparing them mostly based on peak performance and power. The fact that all of the information in the papers comes from public sources can be problematic because some businesses would rather keep their power and performance statistics confidential. Reuther conducts daily news and citation searches to stay current on industry presentations, corporate announcements, and fresh technical press pieces.The fact that five or ten more businesses receive funding, are publicized, and then launch new AI accelerators every year still astounds me, according to Reuther. “When a new group of creative accelerators is introduced, one may believe that the market is sufficiently saturated.
Each study examines a new facet of the subject and summarizes the performance vs. peak power of the existing accelerators. For instance, the 2022 study examined the causes of performance gains and discovered that they result from the use of reduced numerical accuracy (i.e., computing fewer significant digits) and smaller, denser transistor designs. The most recent study looked at several architectural options and analyzed how the system would alter if specific elements were added, such as more cores per processor or parallel performance.
Six new startups have announced their first AI accelerators in the last few months, according to Reuther, who intends to continue the study for the foreseeable future. Lincoln Laboratory should be an objective technical counselor for selecting and developing the appropriate technologies because artificial intelligence (AI) and the hardware that powers it are such hot topics, according to Reuther. “Many sponsors and government colleagues have benefited from our AI accelerator surveys by better understanding the AI accelerator landscape and making more informed research and acquisition decisions. In order to assist all LLSC users as well as our sponsors and mission programs, this poll has been extremely helpful in determining which GPUs we should take into consideration for future LLSC system acquisitions.