CUDA Backend¶
Build-side instructions: GPU Build in the development workflow.
Note
The CUDA/GPU backend is controlled by the SPARQ_ENABLE_CUDA switch (OFF by default; only the CPU
path is compiled). Enabling it requires a local CUDA toolchain (CUDA 13 / CCCL 3 and CUDA 12.8 /
sm_80 on an 8×A800 Linux server tested; verification record: docs/verification/v1c-gpu-cuda-a800.md
in the repository). GPU kernels for the production CondRot primitives are not yet available; the
relevant operators automatically fall back to the CPU path on GPU states.
For large sparse-state runs the per-amplitude payload can be tuned with
-DCACHED_REGISTER_SIZE=<n> (default 64; must cover the circuit’s live
register count, asserted <= 64 on CUDA). The 2^20-dimensional QDA
linear-solver production benchmark runs with 8; combined with the
element-parallel Hadamard_Int_Full kernels this brings the
40568-step walk to ~9.6 GB peak memory and ~1.3 s/step on one A800
(docs/verification/v1d-qda-linsolver-a800.md). The QRAM-free CKS
tridiagonal benchmark runs with 16 (covering the CKS walk’s ~11 peak
simultaneous registers) and completes the 2^20 Poiseuille case in 2682 s
on one A800 (docs/verification/v1e-cks-direct-oracle-linsolver.md).