Benchmarks vs LAPACK / cuSOLVER

[scttfrdmn]

Three-baseline comparison (NKI vs trnsolver-PyTorch vs scipy.linalg/torch on CPU+CUDA) for eigh, eigh_generalized, cholesky, solve_spd. Publish results in docs/benchmarks.md, mirroring the trnfft format. Run on trn1.2xlarge.

[scttfrdmn]

Progress in 968db0b + ac0548c:

CPU half (shipped)

• scipy.linalg baselines (eigh, cholesky, cho_solve) alongside torch.linalg in benchmarks/bench_solver.py
• Real CPU numbers in docs/benchmarks.md
• inv_sqrt_spd_ns (Newton-Schulz all-GEMM variant) added for the Newton-Schulz inv_sqrt_spd via trnblas GEMM #14 track

GPU half (infra shipped; numbers pending)

• benchmarks/bench_cuda.py — torch.linalg on CUDA (cuSOLVER / cuBLAS), @pytest.mark.cuda, auto-skipped without GPU
• infra/terraform/gpu.tf — opt-in g5.xlarge (A10G, Ampere, trn1 vintage peer). Disabled by default (enable_gpu_ci = false)
• scripts/run_cuda_tests.sh — SSM runner mirroring run_neuron_tests.sh
• Vintage-matching methodology documented: trn1↔A10G, trn2↔H100

To populate GPU numbers

cd infra/terraform
AWS_PROFILE=aws terraform apply -var=enable_gpu_ci=true \\
  -var=vpc_id=... -var=subnet_id=...
AWS_PROFILE=aws ./scripts/run_cuda_tests.sh g5

Keeping this issue open until the GPU numbers table in benchmarks.md is populated from a real g5 run.