When people in tech talk about “fast” AI, they usually mean milliseconds—the time it takes a self-driving car to decide whether the object ahead is a pedestrian or a plastic bag. At the Large Hadron Collider, “milliseconds” would be a luxury.
The LHC is the world’s biggest particle accelerator, and experiments such as CMS see roughly 1 billion proton-proton collisions every second. This is far more data than we can store and analyze. Our experiment has only microseconds to decide if a collision is scientifically interesting or just another background event that can be safely discarded.
In the past, CMS relied solely on manually programmed trigger systems to rapidly reconstruct the collisions and separate potentially groundbreaking physics from the overwhelming background (the Higgs boson wheat from the QCD chaff, so to speak.) Our experiment’s custom trigger software was extraordinarily successful, but came with an obvious limitation: we could only search for what we explicitly programmed as “interesting.” (And, as you might imagine, getting 4,000 physicists to agree on the definition of “interesting” is no small feat.)
About 9 years ago, my colleagues and I began asking a simple question: could we build AI tools that are fast enough to help us make these decisions?
The opportunity was enormous. AI could complement traditional trigger algorithms and help us become sensitive to unusual signatures that might not fit neatly into existing theoretical expectations. But the technical challenges were equally daunting.
Every collision produces hundreds of particles that leave signals throughout the detector. Reconstructing what happened is essentially a giant three-dimensional connect-the-dots puzzle. AI excels at finding patterns in messy data, but we cannot simply send these puzzles to a remote computing center for analysis: we don’t have time. (Even though the data travels through fiber optic cables at the speed of light, the unnecessary travel time adds up.) The decision must be made within microseconds, which means the processing needs to happen physically close to our experiment. Also, most AI data farms use a mix of Graphic Processing Units (GPUs) and Central Processing Units (CPUs.) These kinds of computer chips excel at processing large volumes of data, but struggle with ultra-low latency.
That is why we decided to deploy our AI neural networks directly onto field-programmable gate arrays (FPGAs). These are highly efficient and reconfigurable chips that can process data and return a decision with minimal delay. Also, unlike the versatile (but slow) GPUs and CPUs, we can configure our FPGAs to do just one thing very fast.
But then came the biggest challenge: How do you squeeze a powerful neural network onto an FPGA with the smallest area and latency possible?
One way to think about a neural network is as a language. (In this analogy, the FPGA is the “brain,” and the neural networks are how this brain thinks and communicates.) Most neural networks “speak” with tremendous nuance, using 32-bit floating-point precision to represent subtle distinctions. That flexibility comes at a computational cost we cannot afford. Instead, we teach our models to think and communicate in a much more compact dialect—precise enough to capture the important information, but stripped of unnecessary detail. This process is called quantization.
Next, we remove parts of the network that contribute almost nothing to the final decision. Many connections carry weights that are effectively zero. Pruning is like taking a red pen to a dictionary and eliminating words that are almost never used. The model is then retrained to express the same ideas with a smaller, more efficient vocabulary. On an FPGA, removing a connection means the corresponding circuitry is never built, so the hardware (i.e., the “brain”) never spends time or power on it.
The result is a compact, highly optimized neural network embedded directly in the FPGA hardware, yet still reprogrammable whenever we develop a better design.
For CMS, these techniques allow us to isolate the rarest and most scientifically valuable collisions while staying within strict bandwidth and computing constraints. But the benefits could extend far beyond particle physics. These techniques are relevant anywhere real-time decisions matter: autonomous vehicles, fusion-energy control systems, multi-messenger astronomy, wearable medical devices, and neuroscience.
By pushing AI to operate reliably on microsecond timescales while handling extraordinarily complex data, we are doing more than solving a niche problem in particle physics. We are helping define what is possible for fast, efficient AI across science and industry.
