System: Ubuntu 24.04, kernel 6.8.0-147-generic, NVIDIA open driver/GSP 615.71.09, CUDA toolkit 13.3; driver API reports 13.4.
Hardware: Quadro RTX 8000 48 GiB, server variant with stock 250 W limit, originally passive and now actively cooled. An RTX A5000 24 GiB is installed in the same system.
A minimal CUDA Driver API program reproduces elevated idle power on the RTX 8000:
• CUDA initialization/enumeration only: approximately 30 W/P8.
• Retain an RTX 8000 primary context, then block on stdin without allocating a user buffer or launching kernels: approximately 122 W/P0, 0% utilization, 1230 MHz graphics and 6500 MHz memory.
• Release the context while keeping the same process alive: approximately 30 W/P8.
The context-only case uses approximately 167 MiB total VRAM. The behavior reproduces natively and inside a container. The A5000 reaches P8 with its context retained.
Additional controls covered allocations up to 19 GiB, freeing allocations while retaining the context, verified kernel execution and retained CUDA graphs. None is necessary to trigger the RTX 8000 condition. No NVIDIA/Xid journal entries appeared during
testing.
Measurements used selective NVML queries and cumulative GPU energy counters, avoiding continuous full nvidia-smi polling.
Is this expected power-management behavior for this board, or a driver/GSP issue? Is there a supported way to reduce idle consumption while retaining CUDA allocations and preserving inference performance?
nvidia-bug-report.log.gz (974.1 KB)
rtx8000-cuda-context-evidence.zip (5.6 KB)