The system is a research tool that evolves candidate formulas for the Navier-Stokes fluid-flow equations. It hands the surviving candidates to a Lean 4 proof step for machine certification. Search cost is tracked in island-generations. This note covers the hardware under that search: which device did the work, and how a device that did none of it still hung a run.
Why did the RTX 3070 sit idle for the entire project?
The machine runs Windows, and JAX needs WSL2 to use a GPU on Windows. The project's searches ran JAX CPU-only. As a result, the RTX 3070 in the box saw 0 utilization from the search for the whole project.
The record attributes this to the platform requirement. It does not describe any evaluation of the GPU for this workload. This note therefore makes no claim about how the search would have performed on the GPU.
What did running the search CPU-only cost per island-generation?
Each island-generation took about 1.75 seconds, and every search in the project ran at that rate. There is no GPU measurement from this machine, so there is no speedup figure to quote.
The per-generation cost is the term every search budget multiplies: islands × generations × seconds per island-generation. When a platform constraint sets that term, rather than a comparison against alternatives, it is worth recording as a constraint. Otherwise someone may later read it as the ceiling of the hardware.
How can a GPU that is never used hang a job?
One long background run zombied. The recorded cause was JAX hanging while probing CUDA at import, on the machine that held the RTX 3070.
That is the central point of this note. The GPU added nothing to throughput, yet a CUDA probe at import was enough to hang a run. Idle hardware is not the same as hardware the software ignores. A library that checks for accelerators when it loads can interact with the driver stack whether or not it ever computes on the device.
In general, a hang at import is an awkward failure for unattended jobs. Depending on how a job is supervised, it may not surface the way a crash with an error would. That is general reasoning about this class of failure. It is not a record of what this particular run did or did not emit.
How was the CUDA-probe hang at import fixed?
The fix was forcing JAX_PLATFORMS=cpu. That environment variable restricts which backends JAX will initialise, and setting it to cpu limits JAX to the CPU backend. The searches were already running CPU-only, so this pinned the backend the work was using anyway. JAX was no longer left to probe for CUDA at startup.
What should you check before a long JAX run on a Windows machine with an NVIDIA GPU?
These are recommendations drawn from the incident. They are not a list of what this project did beyond the fix above.
- Decide the backend and state it in the environment. If the job will run on CPU, set
JAX_PLATFORMS=cpuinstead of relying on autodetection to reach the same conclusion. - Treat GPU on Windows as a WSL2 question first. Before planning around a card in the box, confirm that the code will run in an environment where JAX can use it.
- Monitor background jobs for progress, not just for a live process. A long job that has logged nothing past import deserves a look.
- Record throughput alongside the device that produced it. A per-generation figure can only be interpreted next to the hardware and backend that set it.
Why does the hardware layer matter for automated search pipelines?
Pipelines like this one are usually judged by what they find, and the compute underneath is taken for granted. Here, the platform set the per-generation cost of every search. A device that did no computation was also behind a hung run. The GPU was in the box the whole time, and the job that hung was hung because of it.