14  FAQ

15 FAQ

15.1 Do I need a GPU to use didgpu?

No. The pure-R and Rcpp backends are full citizens; everything in this book works on a laptop with no GPU. The CUDA backend is optional and detected at install time.

15.2 My CUDA build returns all zeros — what gives?

You’re almost certainly on a Blackwell-class card (RTX PRO Blackwell, RTX 50-series) and an older didgpu (< 0.1.1). Upgrade to didgpu ≥ 0.1.1 — those versions added sm_120 and a compute_120 PTX target.

15.3 How long does a typical bootstrap run take?

For a panel of 200 units × 12 periods × 500 bootstrap reps:

• backend = "r" — minutes to ~30 min depending on estimator;
• backend = "cuda" (Ampere or newer) — seconds to a couple of minutes.

For larger panels the gap widens. See Chapter 8 — GPU & Performance for indicative numbers.

15.4 My bootstrap crashed at hour 11 of 12. Do I have to start over?

No. As long as you passed a checkpoint_dir and a seed, just rerun the same call. didgpu picks up at the first missing cell. See Chapter 10 — Reproducibility & checkpointing.

15.5 Does didgpu agree with DIDmultiplegtDYN / did / fect?

Yes — by design and by continuous test. didgpu’s CI runs a randomized differential-test suite of tens of thousands of panels against each reference package on every push. Current status: bit-for-bit equivalence on the deterministic point estimates and within Monte Carlo error on the bootstrap SEs (which use independent RNG streams).

The few intentional differences (always cosmetic, never inferential) are documented in NEWS.md.

15.6 Can I use didgpu with weighted panels?

Yes. Every estimator family accepts a weight = "wt_col" argument. The four sample-size columns (N, Switchers, N.w, Switchers.w) report both unweighted counts and weighted sums explicitly, so you know exactly what you’re being shown.

15.7 How do I cite didgpu?

citation("didgpu"). didgpu also reminds you in its print output to cite the method paper for whichever estimator you ran (de Chaisemartin & D’Haultfoeuille, Callaway & Sant’Anna, Liu et al., Athey et al., as appropriate).

15.8 Where do I report bugs?

https://github.com/JoshuaAmmons/didgpu/issues. Please include a reproducible example using didgpu_simulate_panel() or your own small panel; the package’s print methods and didgpu_glance() / didgpu_tidy() are useful for paste-into-issue summaries.