Use the GPUs

TraceMaker runs without a GPU and gives the same results with one: the GPU computes cost-to-go fields (and, in the placer, the density and annealing work) faster, but never decides what is legal (decision D5). Every kernel has a CPU reference and a test that both agree bit for bit (GPU acceleration).

Check what it sees

build/release/src/app/tracemaker gpu-info

It lists each device with its name, UUID and the memory it may use. A cpu-only build prints This build has no GPU support (cpu-only). CPU reference paths are used.

Choose devices

The router uses the GPU when the build has CUDA and a device has memory free. --no-gpu computes everything on the CPU. The placer is the other way round: its --gpu option turns the GPU on.

Choose devices by UUID, never by index: CUDA orders the fastest device first (the V100 is device 0 on the development machine), while nvidia-smi lists the P100 first.

nvidia-smi -L                                         # names and UUIDs
CUDA_VISIBLE_DEVICES=GPU-xxxxxxxx-... build/release/src/app/tracemaker route board.kicad_pcb -o out.kicad_pcb

Shared GPUs

The development machine's GPUs are shared with other jobs (a llama-server has held nearly all their memory). TraceMaker treats GPU memory as scarce: a failed allocation or kernel falls back to the CPU with the same result, and the route summary counts the fallbacks (fields 0 GPU + 34 CPU ... 0 GPU fallbacks). Tests must pass with no GPU available; do not stop other people's jobs to free memory.

Building for CUDA

CUDA 12.x only (CUDA 13 dropped Pascal and Volta). nvcc 12.4 rejects GCC 14 and 15 as host compiler, so CUDA host code is built with g++-13 or g++-12 while the rest uses GCC 15 (Build and test).