Arnas View
2D deconvolution inside the camera loop's frame budget.
Arnas View is a 2D deconvolution engine that runs inside the frame budget of a live camera loop. A single-pass restoration of a 4-megapixel frame completes in 2.5 ms, including both transfers to and from the GPU — 395 frames per second, with a 99th-percentile latency of 3.0 ms.
While Arnas Scope (3D) is built to ingest massive volumetric files on workstation storage, Arnas View is built around measured frame latency: it is designed to fit inside the frame budget of a live camera loop, not to process files from disk.
Architecture
The Arnas Architecture: Dynamic & Asynchronous
Arnas View is a 2D camera deconvolution pipeline engineered to fit inside a live camera loop's frame budget. The point-spread function is generated from your instrument's optical parameters and can be regenerated whenever the configuration changes, without a physical calibration step.
Theoretical PSF initialization
The system initializes using an idealized theoretical Point Spread Function — guaranteeing an immediate start without requiring physical calibration phantoms.
Camera noise model
A variance-stabilized restoration mode takes your camera's gain and read noise as inputs, so photon-limited frames are restored against your sensor's actual noise rather than a generic assumption.
Direct GPU acceleration
FFT loops and iterative restoration execute directly on the GPU, maintaining frame-budget latency.
Expanding to 3D deconvolution
Foundation for the full optical toolkit
This 2D architecture serves as the foundation for our broader optical toolkit, naturally complementing our 3D deconvolution engine, Arnas Scope. By shifting complex array operations away from the CPU and handling the transform loops natively via hardware acceleration, Arnas Technologies ensures that both 2D camera frames and deep spatial volumes are processed with speed and mathematical rigor.
Pipeline
Zero-Copy Asynchronous Pipeline
Most GPU-based deconvolution software is built for offline processing: the CPU loads a file, allocates GPU memory, performs the math, and copies the result back. This host-device synchronization adds latency to every frame.
Arnas View bypasses this bottleneck entirely. It is built natively in C++ and CUDA, employing a dual-stream asynchronous pipeline with zero-copy Direct DMA memory transfers.
Pinned memory staging
Input frames are staged in pinned memory, enabling direct GPU access without CPU-mediated copies.
Dual-stream async
The GPU pulls frames via DMA while simultaneously executing R2C and C2R FFTs on the previous frame.
Non-blocking CPU
The CPU is never blocked waiting for the GPU to finish computation — acquisition and processing run in parallel.
Performance
Measured Latency Benchmarks
We do not report "host-enqueue latency" as framerate. Our benchmarks measure the true wall-clock time required to process a frame and fully synchronize it back to host memory — ensuring the frame is actually ready for the OEM camera software to display.
Hardware: NVIDIA RTX 4090, CUDA 12.8
| Mode | Resolution | Mean latency | p95 latency | FPS |
|---|---|---|---|---|
| Single-pass | 2048 × 2048 | 2.53 ms | 2.79 ms | 395 |
| Iterative (2 iter) | 2048 × 2048 | 3.25 ms | 3.61 ms | 308 |
| Single-pass | 4000 × 3000 | 8.65 ms | 9.02 ms | 116 |
| Iterative (2 iter) | 4000 × 3000 | 13.31 ms | 14.15 ms | 75 |
Single-pass restoration of a 4-megapixel frame: 2.5 ms mean, 3.0 ms p99, including both GPU transfers. Timings exclude camera capture and display. A 12-megapixel frame completes in 8.6 ms, inside the 10 ms budget a live loop needs. Hardware: NVIDIA RTX 4090, CUDA 12.8.
Integration
Direct Integration
Arnas View is compiled as a lightweight, C-ABI native DLL (decon2d.dll). It deliberately does not bundle the CUDA runtime or a Python environment, ensuring a minimal footprint that slots directly into existing C++, C#, or Rust camera-control pipelines.
By putting the mathematical rigor of Arnas Scope into a latency-measured SDK, Arnas View allows hardware manufacturers to offer GPU-accelerated optical sectioning and background haze removal directly in their live preview window, within the frame budget of standard scientific CMOS acquisition rates.
Talk about an integrationC-ABI DLL
Callable from C++, C#, Rust, Python — any language with an FFI.
Minimal footprint
No bundled CUDA Toolkit or Python runtime. One driver requirement.
NVIDIA driver only
Requires NVIDIA driver ≥ 566.36 and nothing else on the target machine.
Air-gap ready
Offline activation with hardware fingerprinting for secure facilities.
Common questions
Frequently asked questions
Get started
Ready to integrate high-speed, low-latency deconvolution?
Contact us to discuss your integration requirements, request SDK documentation, or arrange a technical evaluation.
Release notes — see what changed in every version of Arnas Scope.