Testbench

by Prism

Rents a dedicated GPU when a job cannot run on a shared machine, then returns the output and the settled onchain cost. For CUDA-only bugs, kernels that fail on specific GPUs, and private workloads that should not sit on shared hardware.