Remote CUDA Real-Data Acceptance Record

Date: 2026-08-15

Hardware: two 96 GB RTX PRO 6000 Blackwell GPUs

Environment: clean Python 3.12 Conda environment created from environment-remote-cuda.yml, with the current branch installed as a wheel. The environment contained quantem.gpu[cuda,remote] and did not contain quantem.live, quantem.widget, or a separate web frontend.

Exact real-data results

All cold times include HDF5 reading, GPU decode, resident-volume creation, and the first BF response through the HTTP protocol. Warm times are subsequent responses from the CUDA-resident volume. No crop or binning was used unless the row says so.

Evidence	Plan	Cold load	Resident bytes	Exact parity
Seven 512x512x192x192 uint16 tilts	one full tilt per request	1.998 s median	19,327,352,832 each	selected DP and custom detector, max error 0
512x512x192x192 uint32 source	full native detector	4.113 s	19,327,352,832 after lossless narrowing	selected DP, max error 0
1024x1024x192x192 uint16 source	detector bin 2	6.156 s	38,654,705,664	selected DP, max error 0
same 1024 source	detector bin 2, scan bin 2	4.819 s	9,663,676,416	selected DP, max error 0
same 1024 source	direct 256x256 scan crop	0.155 s	4,831,838,208	selected DP, max error 0

The seven-tilt cold range was 1.528-4.136 seconds. With a bounded two-entry cache on each GPU, all seven tilts loaded without exceeding 38,654,705,664 resident bytes per GPU. The least-populated-device policy distributed complete volumes across both GPUs; it did not split a volume or count aggregate memory as single-volume capacity.

Representative warm responses were BF 1.78 ms, ABF 16.68 ms, ADF 7.48 ms, CoM row 18.61 ms, CoM column 2.71 ms, selected diffraction 1.41 ms, and custom detector 3.63 ms. These are service response times, not display-frame-rate claims.

The unbinned 1024 source was rejected in 1.7 ms before allocation: its exact transition peak was 72.0 GB, above the configured 52.2 GB per-device budget. The cropped plan read only the requested scan region. The scan-binned plan replaced the source representation, and changing plans evicted the stale same-file representation.

Failures found and retained fixes

1. An unpinned TestPyPI install selected a build that predates the remote service and therefore had no quantem-gpu command. The repository environment now installs the current checkout. Publish a new release before distributing an environment that installs only from a package index.

2. Admission assumed two bytes per unbinned value. Catalog inspection now reports source dtype and reserves the native item size, including the safe uint32 fallback, before CUDA allocation.

3. File monitoring guessed conventional shard names. It now follows actual HDF5 external links, including absolute paths and nonstandard names. The 1024 evidence correctly reports 2,799,459,964 on-disk bytes instead of only the 7,424-byte master file.

4. CuPy scan binning imported Torch even though it did not use Torch. The import is now optional, and a minimal-environment regression test covers the CuPy path.

Native application gate

The packaged arm64 Live4DSTEM app connected through an SSH alias, launched the service from the standard Conda environment path, cataloged all seven tilts, and displayed non-black diffraction, BF, ADF, thumbnails, histograms, and metadata. Dataset and product switching were exercised through native UI controls on the target display. The structural bundle verifier passed; the full Swift suite passed 39 tests with one opt-in test skipped, and the real SSH/CUDA test passed separately.

This record contains no private hostnames, filesystem paths, credentials, or microscope filenames.