AI hardware is no longer a narrow technical concern. As artificial intelligence reshapes industries and national security, the systems that run AI workloads have become part of the larger question of technological advantage.
That is the problem the Lincoln Laboratory Supercomputing Center has been tracking through the Lincoln AI Computing Survey, known as LAICS and pronounced "lace." Since 2018, the survey has followed commercial AI accelerators, comparing their peak performance and peak power across a rapidly changing market.
Why AI accelerators matter beyond AI
AI accelerators are specialized systems built to speed up work such as neural networks, deep learning, and machine learning. They have been a major focus of hardware development for nearly a decade, and the field has grown quickly enough to require sustained, structured tracking.
Their value is not limited to machine learning. According to the source article, accelerators can also support other parallel applications, including modeling the functions of molecules and speeding up simulations of fluid dynamics. Those are computationally expensive tasks, which makes performance and power consumption central questions for researchers and buyers.
For organizations choosing hardware, the issue is not simply which accelerator is fastest. Different systems are designed for different needs, and efficiency can vary depending on architecture and workload. That makes comparison difficult unless the same basic metrics are gathered and evaluated consistently.
The hardware landscape is broad
LAICS looks across several forms of AI accelerator technology. The source identifies central processing units (CPUs), graphics processing units (GPUs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), and dataflow accelerators as part of the landscape.
Each category comes with different tradeoffs. CPUs can serve general-purpose computing needs. ASICs are built for very specific tasks. Dataflow accelerators, FPGAs, and GPUs offer more flexibility and can be configured for a variety of workloads.
Those distinctions matter because no single accelerator type is automatically best for every use. LAICS is designed to survey technologies currently on the market and compare them so that users can identify the best accelerators for particular needs.
The survey team is led by Albert Reuther and includes LLSC members Michael Jones, Peter Michaleas, Jeremy Kepner, and Vijay Gadepally. The team also works with researchers across Lincoln Laboratory, including in the Advanced Technology Division and the Intelligence, Surveillance, and Reconnaissance and Tactical Systems Division, to understand how accelerators support research and development for their missions.
How LAICS compares the market
The first paper in the LAICS series studied 57 accelerators. The latest paper examined more than 120 accelerators. That growth alone shows why a recurring survey is useful: the number of systems to evaluate has more than doubled across the series described in the source.
The main metrics are peak performance and power. The team also sorts accelerators by whether they are on a chip, card, or system. That gives readers a clearer way to compare systems that may otherwise be described in different formats by different companies.
All of the data in the papers come from public sources. That creates a practical challenge, because some companies prefer to keep performance and power data private. To stay current, Reuther runs daily news and citation searches that surface new technical press articles, company announcements, and industry presentations.
The pace of new entrants remains notable. Reuther says it continues to surprise him that each year another five to 10 startups get funded and announced, then release new AI accelerators. He also plans to continue the survey for the foreseeable future, with six new startups having announced their first AI accelerators in just the past few months.
What the survey reveals about progress
Each LAICS paper does more than list current accelerators. The series also examines a new aspect of the field in each paper, giving readers a way to understand not only what is available, but why performance is changing.
The paper published in 2022 investigated sources of performance increases. It found that gains came from smaller, denser transistor designs and from using lower numerical precision, which means calculating fewer significant digits.
The latest paper looked at architectural choices. It analyzed how adding certain components, such as more cores per processor or parallel performance, would change the system. That focus is important because accelerator design is not a single path; different architectural decisions can shape the balance between performance, power, flexibility, and suitability for a given workload.
Why unbiased comparison is valuable
The Lincoln Laboratory Supercomputing Center operates and optimizes high-performance computing systems used by thousands of laboratory research staff. For that audience, accelerator choices affect not just individual projects, but the broader computing environment that supports research work.
Reuther describes Lincoln Laboratory’s role as an unbiased technical advisor for choosing and pursuing the right technologies. The surveys have helped sponsors and government colleagues better understand the AI accelerator landscape and make research and acquisition decisions.
The work also feeds back into LLSC’s own planning. The survey has been valuable in deciding which GPUs to consider for upcoming LLSC system purchases. In that sense, LAICS supports both external sponsors and the users who rely on the center’s computing systems.
The broader takeaway is straightforward: AI accelerator hardware is evolving quickly, and buyers cannot rely on simple labels or isolated performance claims. Comparing peak performance, power, form factor, and architecture from public data gives decision-makers a clearer basis for choosing systems that fit real workloads.