GPU environment of the benchmark machine

State of the machine used for docs/perf.md, recorded on 2026-10-22, and what it takes to run QoraPhase on its GPU. Everything under "Measured" was read from this machine; the rest is marked as such.

Measured

item	value
GPU	NVIDIA GeForce RTX 5070
compute capability	12.0 (nvidia-smi --query-gpu=compute_cap)
VRAM	12,227 MiB total (11.6 GB free of 12.8 GB reported by CuPy mem_info with the desktop running)
driver	610.88, WDDM driver model, reports CUDA 13.3
CPU / RAM	AMD Ryzen 5 7500F (6 cores, 12 threads), 31.6 GB
OS	Windows 11 Home 10.0.26300
Python / uv	3.14.5 / 0.11.32
CUDA toolkit	not installed (nvcc not found)
WSL	installed, but no distribution
JAX in the project env	0.11.1 (jaxlib 0.11.1), CPU backend only, jax.devices() is [CpuDevice(id=0)]

Can JAX use this GPU?

Not on native Windows. uv pip install --dry-run jax-cuda13-plugin and jax-cuda12-plugin find no installable wheel; the resolver lists only manylinux_2_27_x86_64 and manylinux_2_27_aarch64 builds (v0.12.0.1). So the options for JAX on the GPU are:

1. WSL2 with a Linux distribution and jax[cuda13] inside it. Not tried: it needs a distribution installed (wsl --install, which may require an administrator prompt and a reboot) and the Windows NVIDIA driver, which is present. Whether the RTX 5070 (compute capability 12.0) works with the JAX CUDA wheels under WSL2 has not been tested.
2. A Linux machine with an NVIDIA GPU.

Blackwell (sm_120) support in the JAX CUDA wheels

The RTX 5070 is compute capability 12.0 (Blackwell). Inspected without running anything: the Linux wheels jax-cuda13-plugin 0.12.0.1 and jax-cuda13-pjrt 0.12.0.1 were downloaded (pip download --platform manylinux_2_27_x86_64) and their binaries searched for target strings. jax_plugins/xla_cuda13/xla_cuda_plugin.so (396 MB) contains sm_120 (12 times), sm_121 (9), sm_100 (19) and compute_100 (31), and no compute_120; the plugin extension modules contain none. That shows the XLA backend knows these architectures; it does not show that a program runs on this card, and the CUDA 13 line of wheels (as opposed to CUDA 12) is the one that targets Blackwell. The with-cuda extra pulls nvidia-cuda-nvcc 13.x (for ptxas). Whether JAX 0.12 runs on the RTX 5070 under WSL2 is untested here; the project env is JAX 0.11.1 and would have to be upgraded to 0.12 to use these wheels.

For any VRAM measurement set XLA_PYTHON_CLIENT_PREALLOCATE=false; otherwise JAX reserves 75% of the device memory at start-up and the peak-memory figure reflects the reservation, not the use.

What does work today: CuPy on native Windows

A throwaway environment outside the repository (uv pip install "cupy-cuda13x[ctk]", CuPy 14.2.0, CUDA runtime 13.2 from wheels, no toolkit install) sees the GPU and runs cuFFT:

• device name NVIDIA GeForce RTX 5070, compute capability 120, runtime 13020, driver 13030;
• cp.fft.fftn of a 256 x 256 x 256 complex64 array took 1.6 ms per call (mean of 10 calls after one warm-up, synchronized), and ifftn(fftn(a)) returned a to a maximum absolute error of 7.8e-7.

This is one probe, not a benchmark: one operation, one size, no comparison with the CPU run under identical conditions. For scale only, the CPU run in docs/perf.md measured 73 ms for fftn + ifftn at 256^3 on JAX (about 36 ms per transform if the two cost the same).

CuPy is not a dependency of QoraPhase and was not added to the project.

Consequences for the GPU track

• GPU numbers for the existing JAX code require option 1 or 2 above. Until then, no GPU figure for QoraPhase exists; docs/perf.md stays CPU-only.
• A CuPy port of the RAAR/ER loop would run natively on this machine, but it would be a second implementation. It is only a fair comparison against the JAX code if the two are shown to produce the same results to a stated tolerance, and the issue asks for equal algorithms and work.
• A custom CUDA kernel (QoraPhaseCU) additionally needs the CUDA toolkit (nvcc) and a compiler supported by it; neither is installed. cupy-cuda13x[ctk] ships the runtime and libraries but not nvcc.