IN Brief:
- Cambridge-based Signaloid has joined CERN openlab and will deploy its UxHw technology in the Heterogeneous Architectures Testbed.
- CERN will test the architecture on Monte Carlo particle-event generation, measuring performance, numerical accuracy, and integration effort.
- The work follows Signaloid's move from software and FPGA implementations towards dedicated ASIC and chiplet hardware.
Signaloid has joined CERN openlab as an industrial member and will deploy its UxHw distribution extended computing technology within the laboratory’s Heterogeneous Architectures Testbed. The joint programme will examine whether the architecture can accelerate Monte Carlo event generation for the High-Luminosity Large Hadron Collider while preserving numerical accuracy and fitting existing scientific software workflows.
The Cambridge company has developed UxHw for calculations involving uncertain or probabilistic inputs. Conventional Monte Carlo methods repeatedly execute a calculation using different sampled values and then analyse the resulting distribution. Signaloid’s architecture instead operates directly on digital representations of probability distributions, potentially reducing the number of repeated executions required for suitable workloads.
CERN and Signaloid will evaluate a representative particle-event generation workflow based on the Pepper framework, covering proton-proton collisions that produce multiple gluons. The programme will measure computational performance, numerical accuracy, and the amount of integration work needed to place the technology into an established high-energy physics software environment.
Those measures are closely connected. A specialised accelerator can produce a strong standalone benchmark while delivering less benefit once data movement, software adaptation, runtime scheduling, and numerical verification are added around it. CERN’s testbed provides an environment in which the architecture can be compared with existing CPU and GPU infrastructure as part of a wider heterogeneous computing system.
The High-Luminosity LHC increases the pressure behind that work. CERN expects computing requirements to rise as the upgraded collider produces substantially more recorded collisions, increasing the amount of simulation and analysis required alongside experimental data. Monte Carlo event generation already consumes a significant share of computing resources because large numbers of simulated collisions are needed to compare theoretical models with physical measurements.
Signaloid reports speed-ups of as much as 2,000 times on selected representative workloads, but those figures remain company benchmarks and do not represent a universal processor comparison. Workloads with limited stochastic calculation would not be expected to gain the same advantage, while any result at CERN will also depend on the software pathway surrounding the accelerated computation.
The test comes as Signaloid moves UxHw further into dedicated silicon. The company has already taped out a custom ASIC on a low-power TSMC process and recently extended the architecture towards chiplet implementations. That progression creates several possible deployment models, from software and FPGA acceleration through dedicated ASICs and eventually heterogeneous packages containing specialised UxHw dies.
Chiplets could make the architecture easier to place beside more conventional processing resources when probabilistic computation forms only part of an application. The resulting system would still need efficient die-to-die communication, memory access, and software scheduling to prevent the cost of moving data between processors from eroding the acceleration achieved inside the specialist hardware.
CERN’s Heterogeneous Architectures Testbed is intended to expose precisely those practical boundaries. Rather than replacing CPU and GPU resources wholesale, new processors can be evaluated for particular workload classes and incorporated where they offer a measurable advantage. Distribution extended computing therefore enters the programme as another candidate accelerator rather than as a proposed single architecture for the complete scientific computing stack.
The collaboration also gives Signaloid an external test of a technology whose earlier performance figures have largely come from its own benchmarking. Scientific event generation is particularly useful for such an evaluation because both execution time and numerical behaviour can be measured against established methods. An accelerator that shortens runtime but alters the probability distribution in an unacceptable way would have limited value for physics workloads.
The project is expected to identify which parts of the Monte Carlo pipeline suit UxHw and where conventional processors remain preferable. With High-Luminosity LHC operation expected to begin around 2030, CERN has several years to test emerging architectures before computing demand rises further. Signaloid’s contribution will be judged less by the novelty of processing probability distributions directly than by whether that approach can be integrated, validated, and operated alongside the heterogeneous hardware already supporting scientific computing.
