Name the GPU a benchmark used; add an Arm data point

[jacotay7]

Fix

benchmarks/run.py recorded the GPU as the first line of nvidia-smi, which ignores CUDA_VISIBLE_DEVICES. On cfl-test-bench (RTX A400 and RTX 4060), a run on the RTX 4060 was labelled "NVIDIA RTX A400". The numbers were the 4060's: up to 4× the A400's on the large configurations. The script now asks CuPy for the current device's name. It also reads the CPU model from lscpu on Arm, where /proc/cpuinfo has no model name, so it reports Neoverse-N1 instead of aarch64. Checked on both cards with CUDA_VISIBLE_DEVICES and locally (x86, Quadro P620).

Data

• New artifacts: benchmarks/device-results-neoverse-n1-rtx4060.json (CPU and GPU) and benchmarks/device-results-neoverse-n1-rtxa400.json (GPU). They were re-measured with the fixed script on makewfs 2.0.0, pinned to 16 cores.
• New table: benchmarks/device-results-neoverse-n1.md puts N1 CPU, RTX A400 and RTX 4060 next to the existing Ryzen 9 9950X3D and RTX 5090 numbers.
• Docs: a short paragraph in docs/performance.md.
• On the bench: the representative regression check passes, and the full test suite passes (GPU tests included) except the known OOPAO comparison. The bench's OOPAO checkout is a fork with a different wfs_measure signature.

🤖 Generated with Claude Code