Real-time Analytics
CYRAN needed to deliver interactive geospatial intelligence in air-gapped and sovereign environments, but JPEG 2000 decode latency on large satellite scenes was the bottleneck that kept edge analytics from being practical.
Reported impact
CYRAN cut decode time on a 26 GB satellite scene from more than five minutes to just over two minutes, outperformed a CPU baseline by about 3x and a discrete GPU workstation by up to 2x, and made genuinely interactive geospatial analysis possible on local hardware without cloud dependency.
CYRAN reworked the slowest stage of its SpatialFuse geospatial pipeline around NVIDIA DGX Spark rather than trying to push ever-larger scenes through traditional CPU or discrete-GPU workstations. The company built PixelFlux on top of nvJPEG2000 so compressed satellite imagery can be decoded directly on the GPU, and DGX Spark's unified memory means those buffers are immediately available to the CPU and downstream analytics without PCIe copy overhead. SpatialFuse then uses the decoded scenes for edge inference across orbit, air, and ground data sources in secure environments where cloud round-trips are not acceptable. DALI integration keeps the GPU fed from local storage for training and inference pipelines, and the same architecture now gives CYRAN a path from local edge validation to larger Blackwell-based cloud or data-center deployments without rewriting the core geospatial stack.
Compressed satellite scenes enter PixelFlux, which uses nvJPEG2000 on DGX Spark to decode large JPEG 2000 imagery directly into unified memory. SpatialFuse then fuses those decoded scenes with other sensor inputs and serves geospatial analysis locally in air-gapped or forward-deployed environments.
By combining the unified memory architecture of NVIDIA DGX Spark with GPU-accelerated JPEG 2000 processing through its new PixelFlux library, CYRAN cut the decode time on a 26 GB satellite scene from over five minutes to just over two.
Interactive large-scene analysis runs entirely on local hardware-enabling air-gapped and sovereign intelligence deployments.
On NVIDIA DGX Spark, the GPU and CPU share the same memory pool, so decoded image data is immediately available without any transfer step.
The first was GPU-accelerated JPEG 2000 decode. NVIDIA nvJPEG2000 moves the most computationally demanding stages of image decoding off the CPU and onto the GPU.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Inc42's embedded company schema lists "numberOfEmployees":{"@type":"QuantitativeValue","value":24}, which maps to SMB under the 0-249 rule.
Related
NVIDIA Sample Jun 16, 2026 NeoSpace NeoSpace needed to make tabular foundation models viable for financial institutions that serve tens of millions of customers, but time-to-train, inference latency, and the cost of running transformer workloads on… 10% increase in installment credit offers NVIDIA CUDA libraries · NVIDIA GB200 NVL72 · NVIDIA TensorRT LLM NVIDIA Sample Jan 12, 2026 XPAI XPAI needed a privacy-preserving way to measure how people use and perceive physical spaces so brands and event operators could optimize offline experiences with real-time evidence instead of intuition. millions of data points per second AVAI One Sensor · NVIDIA DeepStream SDK · NVIDIA Jetson Orin 
Google Cloud Sample Archived · Oct 16, 2026 Alta Alta needed a transactable Google Cloud Marketplace path for its AI and technology spend management platform without pulling its product team off core forecasting work, while also helping finance and IT leaders… Saved 125 engineering and 60 project management hours Cloud Pub/Sub · Cloud Run · Cloud SQL 