TDSE Circuit can execute fixed-topology linear transient steps on the CUDA path when the GPU Add-on is installed and authorized. The runtime reports the selected execution path and a stable reason code through Circuit diagnostics. diagnostics.gpu_resident.product_runtime_selected_path uses the TDSE_CIRCUIT_GPU_RUNTIME_PATH_* macros below; the first column lists the macro suffix so the table remains readable.

Path macro suffix	Meaning	Fallback
`FIXED_LINEAR_CUDA_GRAPH`	Device-resident CUDA graph path with no host transfer in the measured step loop.	None.
`FIXED_LINEAR_CUDA_GRAPH_HOST_OUTPUT_BOUNDARY`	CUDA graph path with an explicit final host-output copy.	None.
`GPU_CANDIDATE`	The workload is supported for evaluation but is not a qualified product path.	Structured CPU fallback remains available.
`CPU_FALLBACK`	A resident CUDA-graph run violated the one-launch product invariant.	CPU fallback with the `...REASON_LINEAR_CUDA_GRAPH_LAUNCH_COUNT_MISMATCH` reason macro.
`UNREPORTED`	No resident GPU execution was recorded for this call.	Inspect the normal call status and diagnostics; this is not itself a CPU-fallback classification.

The currently supported GPU product path is dense, fixed-topology, linear transient execution. Sparse CUDA and nonlinear variants remain candidates until their product qualification is complete.

Inspect the tdse_circuit_diagnostics_t diagnostics field returned by the relevant Circuit compute call to obtain the selected path, reason code, selected device, transfer counters, and execution-state diagnostics. Use the SDK's circuit_resident_transient_probe for a local functional check. Performance campaigns, hardware-matched baselines, cloud execution, and release reports are owned by the companion tdse_benchmark and release-infrastructure repositories.

Resident Transient Runtime Policy

On this page