What software makes the GPU usable?
A GPU can be installed correctly while the application cannot use it. Check the software layers separately.
Four different software layers
- Application and framework
The workload's own environment
- Runtime and libraries
The interfaces used during execution
- Host driver
Manages the installed device
- GPU hardware
Must be supported by the stack
Know the layers
The driver manages the device. The CUDA toolkit includes development tools and libraries. Frameworks and application libraries build on those interfaces.
A container packages the application environment. It still depends on a suitable host driver.
The CUDA version reported by a driver tool is not proof that the corresponding compiler is installed.
Inspect without changing
On an authorized NVIDIA Linux host with the tools installed:
nvidia-smi -L
nvidia-smi
nvcc --versionThese check detected GPUs, driver/device information, and the compiler respectively. They are not Apple GPU commands or a complete workload test.
Record the application image and runtime versions too. Change one layer deliberately so you retain a useful comparison and rollback path.
Try it yourself
A prebuilt application works, but the machine has no nvcc command. Is that a contradiction?
Check your answer
No. It may need runtime libraries and a driver without needing a local compiler. Driver support, toolkit installation, and application dependencies are different checks.