Skip to main content
Ctrl+K

quantem.gpu

Start here

  • Install

Python workflows

  • Gold: I/O, DPC and SSB
    • Advanced: probe and reconstruction
  • Save and share your data

API reference

  • Python API
    • Load data and select patterns
    • BF, ADF, and DPC images
    • Direct ptychography and aberrations
      • Choose a direct reconstruction
      • Portable SSB results
    • Drift across a time series
    • Movie API
    • Devices, optics, parallax, and scan rotation

Developer guide

  • Developer guide
    • Scientific operations
      • Scientific contract
      • Data model and coordinates
      • Load, decode, and bin
      • BF, DF, and ADF reductions
      • CoM, DPC, and iDPC
      • Single-sideband ptychography
      • Explicit scan regions
      • Display and export kernels
        • Shared Metal image statistics
    • Backend implementations
      • Python MPS
      • CUDA
      • Native Swift and Metal
        • Native Swift and Metal
      • WebGPU
      • CPU reference
    • Native application integration
      • Load native acquisitions
        • Native acquisition formats and metadata
        • Original HDF5 packed loading
        • Native encoded inputs and scaled output
      • Native Metal image endpoints
      • Native SSB fit
      • Regional mean diffraction from compressed residents
    • File formats and codecs
      • QEM specification
        • QEM codec layouts
        • QEM metadata mapping and completeness
        • QEM references and interoperability checks
        • K3 DM4 and saved .qem copies
      • Dense and lossless-packed data
      • Scaled uint16 resident intensities
      • Prepared packed format
        • Native Lossless Pack Format v1 producer
      • Paired ARINA microscope metadata
      • Count-ANS container
      • Paired-count resident layout
    • QuantEM.GPU Remote
      • Deploy a Linux CUDA host
      • Connect locally or over SSH
      • Protocol and integration
      • GPU admission and residency
    • Testing and benchmarks
      • IO and compression
        • CUDA count codecs: speed and peak memory
        • IO reports
      • SSB
        • Full-field and regional SSB benchmark
        • SSB reports
      • Virtual detectors
        • Virtual-Image Kernel Checklist
      • CoM, DPC and iDPC
      • Rendering and transfer
        • Rendering reports
      • Cross-backend parity
      • Kernel registry
      • Benchmark methodology
      • Testing and evidence
      • Verified benchmark results
      • Revision and change ledger
      • Backend Optimization Matrix
      • Backend 4D-STEM Load/Decode Checklist
      • Strict ANS acquisition acceptance
      • Backend coverage
    • Contributing changes
      • Kernel architecture
      • Repository architecture
      • Adding a backend or kernel
      • Count and Fourier representation integration
      • Reproducing resident count analysis
      • SSB architecture
      • Resident joint SSB fitting
      • Writing, notation, and units
    • Historical records
      • Packed-resident optimization evidence
      • Migration notes
      • Native Metal HDF5 Loader Postmortem
      • M2 Air Metal evidence
      • Remote CUDA evidence
      • Frozen MPS fixtures on a virtual GPU (2026-09-20)
      • SSB units and thick-sample model (2026-09-24)
      • CUDA pool retention (2026-10-04)
      • Package restructure (2026-10-05)
      • Downstream sync after the restructure (2026-10-05)
      • SSB detector sampling and fit state (2026-10-05)
      • Parallax and scan rotation on encoded input (2026-10-05)
      • MPS uint32 masters and the CUDA 13 extra (2026-10-07)
      • uint32 flagged markers and convert size (2026-10-08)
      • WebGPU .qem streamed loading (2026-10-08)
      • LZ4 read bounds and rc16 timing (2026-10-11)
      • Metal LZ4 bounds (2026-10-11)
      • Viewer reductions, float means and gzip stack loads (2026-10-11)
      • Speed round after rc16 (2026-10-11)
      • Releasing cached GPU memory (2026-10-11)
      • Stored SSB fit schema (2026-10-11)
      • Calibration speed-up on main (2026-10-13)
      • Screening and SSB share acquisition storage (2026-10-16)
      • MPS Fourier transforms for iterative waves (2026-10-18)
      • Native detector median contract (2026-10-23)
      • Original-load median fill (2026-10-25)
      • Native direct drag windows (2026-10-25)
      • Native Metal SSB migration
      • SSB performance notes (to 2026-10-18)
      • Historical experiments
        • WebGPU G(q,k) memory history
        • WebGPU frame-cooperative decode
      • Experimental resident ANS
      • Experimental Metal entropy series (2026-09-08)
  • .md

IO reports

IO reports#

Dated records retain their original workload, backend, source revision and acceptance limits. They are evidence, not a statement of current release speed.

  • QEM conversion and complete lossless verification, 2026-10-23

  • QEM index size and CUDA loading, 2026-10-23

  • QEM compression compared with HDF5, 2026-10-23

  • Bounded MPS HDF5 loading, 2026-10-23

  • Acquisition-worker loader preparation, 2026-10-23

  • Native Swift/Metal IO benchmark, 2026-10-23

  • Python MPS IO benchmark, 2026-10-23

  • Disk–GPU loading: HDF5 and QEM, 2026-10-23

  • CUDA IO benchmark, 2026-10-23

  • Six default acquisition IO comparison, 2026-10-23

  • Five-acquisition Disk–GPU loading comparison, 2026-10-23

  • CUDA QEM loading: cached-block scheduling, 2026-10-23

  • CUDA QEM loading and file-cache recheck, 2026-10-23

previous

CUDA count codecs: speed and peak memory

next

QEM conversion and complete lossless verification, 2026-10-23

By Sangjoon Lee et al.

© Copyright 2026, ophusgroup.