VortexSynth rebuilds GPU benchmarks to test new hardware designs
VortexSynth: Profile-guided benchmark reconstruction for GPU architecture research
Hardware ArchitectureProgramming Languages
Summary
It can be hard to test new graphics processing units (GPUs) because many useful programs are closed-off or use software that won't run on new designs. The authors created VortexSynth, a tool that watches which parts of a program are most important, rewrites those parts to work on the new GPU, and checks that the output matches the original. This process repeats to get closer to the performance and behavior of the original program. The tool was tested on an open GPU architecture and showed it can successfully recreate useful workloads.
What this means in practice
- •For gpu hardware designers: Test GPU architectures by reconstructing benchmarks automatically from existing workloads without manual rewriting.
- •For robotics software teams: Evaluate new GPU designs using automatically adapted key kernels from robotics workloads for better hardware-software integration.
Authors
Huanzhi Pu, Chulhyung Park, Chihyo Ahn, Ramyad Hadidi, Hyesoon Kim
Abstract
GPU architecture research requires representative benchmarks that can execute on new designs. However, many vendor-tailored workloads and library kernels cannot run directly on research architectures because their implementations are closed-source or rely on programming models unsupported by the target architecture. Migrating existing kernels requires manual effort to adapt their implementations and validate their behavior. This creates a need to automate workload migration to target platforms. We present VortexSynth, a framework that uses profile-guided benchmark reconstruction to construct GPU benchmarks from reference workloads. Our workflow selects hot kernels through profiling, synthesizes implementations in the target programming language, verifies output equivalence, and measures profile differences using a translated proxy implementation. The resulting profile signature gap is supplied to an agent for iterative refinement. We evaluate the method on Vortex, an open-ISA GPGPU, using robotics as the application domain. All active kernel pairs pass both output-equivalence checks, and the case studies show that iterative refinement can reduce profile distance.