Radiance

Radiance is an open-source GPU platform developed at UC Berkeley. It is designed to help researchers and developers build and evaluate full-stack heterogeneous GPU systems targeting AI and neural graphics workloads.

Radiance provides:

• A scalable integration platform for heterogeneous accelerators, with clear interfaces for decoupled orchestration and data delivery.
• A clean-slate SIMT core designed for modern, register-heavy GPU kernels and optimized for ASIC implementation.
• Chipyard-based SoC integration, enabling tight host-device coordination for memory system behavior and kernel execution.

Radiance is under active development. Please refer to the following resources for more information:

• Radiance GPU architecture
• Muon SIMT core and command processor architecture
• Neutrino orchestration unit architecture

Muon SIMT Core

Muon, the SIMT core design for the Radiance GPU, is currently under active development.

See spec documents:

• ISA specification
• Muon microarchitecture specification
• Memory system specification

Run simulations

Run these commands in ${CHIPYARD_ROOT}/sims/vcs after source env.sh. Use the graphics branch for Chipyard.

Full SoC config with multiple Muon clusters/cores, RTL memory subsystem and a host CPU core:

make CONFIG=RadianceTapeoutSimConfig run-binary LOADMEM=1 BINARY=...

Below simulates a GPU-standalone, core-isolated config, useful for functional verification and fast debugging iterations:

Single Muon core, with Cyclotron-based IMEM and GMEM memory models, with differential-testing enabled against Cyclotron:

make SUB_PROJECT=muon CONFIG=MuonCoreTestConfig run-binary BINARY=...

Single Muon core, with Cyclotron-based IMEM and GMEM memory models, with differential-testing disabled:

make SUB_PROJECT=muon CONFIG=MuonCoreNoDiffTestConfig run-binary BINARY=...