Invited Talk by Thea Aarrestad (ETH)

Tile: Pushing the limits of real-time ML: Nanosecond inference for Physics Discovery at the Large Hadron Collider

Abstract: At the CERN Large Hadron Collider (LHC), high-energy proton collisions generate O(10,000) exabytes of raw data annually. To manage this immense data volume while adhering to computational and storage limitations, real-time event filtering systems must process millions of proton-proton collisions per second, utilising a staged system of FPGAs, CPUs and GPUs to perform efficient reconstruction and decision-making. Within a few microseconds, over 98% of the collision data must be discarded both rapidly and accurately. As the LHC transitions to its high luminosity phase (HL-LHC), these systems - situated in radiation-shielded caverns one hundred meters underground - will confront data rates equivalent to 5% of global internet traffic, combined with unprecedented data complexity.

To maintain data integrity for meaningful physics analyses, highly optimized machine learning (ML) algorithms are being employed for real-time data processing. This demands the development of novel methods and tools to achieve extreme throughput, ultra-low latency, and low-power inference on specialized hardware.

In this presentation, we will discuss how real-time ML techniques are employed to process and filter vast amounts of data, enhancing physics signal acceptance. We will explore state-of-the-art methods for designing and deploying high-speed ML algorithms on FPGAs, ASICs, and GPUs. Lastly, we will examine the applications of real-time inference in particle physics experiments and its critical role in facilitating the discovery of New Physics.