Artificial intelligence is reshaping industry and national security, making it essential to understand the latest hardware capabilities to maintain a technological edge. AI accelerators—specialized systems that speed up neural networks, deep learning, and machine learning—have seen rapid development over the past decade. Since 2018, the Lincoln Laboratory Supercomputing Center (LLSC) has run the Lincoln AI Computing Survey (LAICS), publishing six papers that continuously summarize commercial AI accelerators' peak performance and peak power.
The survey was motivated by a sharp rise, about eight years ago, in research papers describing new AI accelerators and by government sponsors asking for guidance on these technologies. Albert Reuther, the project lead, and his team manage and optimize the high‑performance computing systems used by thousands of laboratory researchers.
Beyond machine learning, AI accelerators enable other compute‑intensive parallel tasks such as molecular function modeling and fluid‑dynamics simulations. Accelerator forms include CPUs, GPUs, ASICs, FPGAs, and data‑flow accelerators. CPUs serve general‑purpose workloads, ASICs perform narrowly defined tasks, while GPUs, FPGAs, and data‑flow designs are more flexible and can be tuned for a variety of workloads. Design choices lead to differing efficiency and performance characteristics.
LAICS compares accelerators using publicly available data on peak performance and power, categorizing them by chip, card, or system level. The first paper examined 57 accelerators; the latest covers more than 120. Most data come from technical reports and company announcements, which can be hard to obtain because some vendors keep performance and power figures private. Reuther runs daily news, citation, and industry‑presentation searches to stay current.
The survey finds that each year roughly five to ten startups receive funding and announce new accelerators, indicating a still‑vibrant market. A 2022 paper traced performance gains to smaller, denser transistor designs and the adoption of lower numerical precision (fewer significant digits). The most recent paper explores architectural choices—such as adding more cores per processor or increasing parallel performance—and how they affect overall system behavior.
Reuther intends to keep the survey going for the foreseeable future; in the past few months six new startups have launched their first AI accelerators. He notes that AI and its hardware are hot topics, and Lincoln Laboratory must remain an unbiased technical advisor for government and research programs. LAICS has helped sponsors and colleagues better understand the accelerator landscape, make smarter research and acquisition decisions, and guide LLSC’s upcoming GPU purchases, benefiting all users.
Review