From one economy to a GPU batch

Synthetic English narration.

One economy is an array
One economy contains a firms-by-markets array. Its independent market problems provide work for the GPU, although the steps inside each problem still depend on earlier steps.

Add independent economies
Place several compatible economies in one batch. The extra dimension identifies the economy. Each keeps its own parameters, market shares, and stopping state.

Give the device more work
The batch exposes more independent work to GPU thread blocks. Several economies can occupy the device together, instead of submitting each complete economy separately.

Keep data available
Keep working arrays on the device and reuse compatible operations. This can reduce repeated setup and transfers. Finished economies can leave the active batch.

Choose a batch that fits
A larger batch also needs more memory. Occupancy, temporary arrays and uneven convergence affect the result. Measure the complete solve; batch size alone does not guarantee speed.