This is a short public record of the Birds of a Feather session on the state of open-source computer architecture simulation, so its discussion and takeaways reach beyond the room. The session saw a panel of five experts present their role in the community, and their thoughts on the open-source architectural simulation landscape, alongside an open discussion with the attending audience members.
What the session set out to do
The BoF put three questions to the open-source simulation community: what do we have, what do we need, and what do we want? The motivation came from the community's own recurring complaints heard consistently across various research venues. Simulators are too slow or too hard to learn, thin documentation, tools that are hard to adapt, a fragmented landscape, and little structural incentive for researchers to build sustainable, validated, well-documented tooling.
The state of open-source tooling was taken through an optimistic perspective, we have capable tools, but the question is how to refine, combine and extend them, on both the user and developer sides.
What the panel brought
- Adolfy Hoisie (Brookhaven National Laboratory) surveyed the well-defined modelling-and-simulation methodologies alongside their trade-offs between speed, accuracy and generalisability, and argued for a new class of open-source, AI-driven simulators built around a clear structure: model development, training sets, interfaces to traditional simulators, verification and validation, extensibility, generality, and interfaces to LLM/agentic AI. What we want: open, AI-driven simulation infrastructure.
- Filippo Spiga (NVIDIA) framed open-source tools as a workable "middle ground" for co-design between academic and industry partners and set out the next challenges: more sophisticated nested hierarchies, treating the full system rather than units in isolation, composition and interoperability of tools, and a common trace language. What we need: interoperability and a shared interchange language.
- Ron Brightwell (Sandia National Laboratories) presented the Structural Simulation Toolkit (SST) and its goal of interoperable components, and, through the AGILE project, named the hard open problem: a full system can be modelled in parts, but passing results between models of different fidelities, and understanding the effect on accuracy, remains unsolved. What we need: multi-fidelity composition we can trust.
- Artur Podobas (KTH Royal Institute of Technology) asked whether we can combine approaches (full-system simulation, RTL simulation, and FPGA-empowered simulation) for better precision, results, or speed. What we need: hybrid, combined simulation approaches.
- Yukinori Sato (Toyohashi University of Technology) offered his views on contending alternatives for the what we... components contained within the BoF title. We have binary instrumentation as an under-used alternative to the usual simulator/emulator, we need a push toward system-level modelling that captures the cache, OS and software stack together rather than the architecture alone, and we want to make simulation fast enough on real applications to carry results through to real-world deployment. He grounded the latter with a concrete example: cycle-accurate simulation of a Gemmini accelerator driven by custom RISC-V instructions. What we want: simulation that reaches real-world deployment.
What the session concluded
The room surfaced many ideas that relate to problems rather than agreed directions of resolution. A few things came through clearly:
- Real progress will be most successful with a consistent communication medium rather than by occasional discussion at, possibly niche, research venues. This may need to be championed by a committed few.
- The most promising directions were asynchronous and concrete, agreeing a standard interoperation (trace) language, or standing up a community-run hub for validating open-source simulation contributions, one credible enough to be cited in publications as evidence for the claims a tool makes.
- Whilst everyone wants this, the effort-in versus value-out payoff is poor until the work is largely complete, and it attracts little academic credit along the way, which is precisely why it needs a coordinated community effort rather than isolated individuals.