Higher Education / Research
CYRAN AI Decodes Gigapixel Satellite Imagery On The Edge With NVIDIA DGX Spark
Objective
CYRAN AI Solutions, an AI-native deep tech startup spun out of IIT Delhi’s NVM and Neuromorphic Hardware Research Group, builds sensor fusion platforms for geospatial intelligence, security, and Industry 4.0 applications. Its SpatialFuse platform fuses data from orbit, air, and ground into real-time geospatial intelligence—but for the analysts and engineers who rely on it at the edge in orbit and in air-gapped environments, the pipeline is only as fast as its slowest step. That step was decoded.
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—delivering truly interactive geospatial analysis on local hardware, without a cloud in sight.
Customer
CYRAN AI Solutions
Partner
Indian Institute Of Technology Delhi (IIT Delhi)
Topic
Edge Computing
Geospatial AI / Edge Inference
Products
Key Takeaways
3x Faster
- JPEG 2000 decode vs. CPU baseline—a 26 GB satellite scene drops from 5+ minutes to just over 2 minutes.
2x Faster
- Than a discrete GPU workstation at large image sizes, by eliminating a hardware-level memory bottleneck that DGX Spark’s architecture eliminates entirely.
Zero Cloud
- Interactive large-scene analysis runs entirely on local hardware—enabling air-gapped and sovereign intelligence deployments.
Decoding a Planet’s Worth of Data
Satellites, drones, and ground sensors are generating geospatial data at a pace that has outrun the tools built to handle it. CYRAN’s SpatialFuse platform is designed to deliver intelligence at the edge—in forward operations centers, air-gapped secure environments, and eventually aboard space platforms—and every stage of the pipeline has to perform. The AI model isn’t the bottleneck. The decode step is.
JPEG 2000 is the dominant compression format across Earth observation—used by Sentinel-2, USGS, Maxar, and virtually every major satellite data source. Single scenes routinely exceed 10 GB uncompressed. On a standard CPU pipeline, decoding a 26 GB satellite scene took over five minutes. That’s not a minor inconvenience. It’s a pipeline stopper that breaks iterative analysis workflows and makes interactive large-scene exploration effectively impossible.
The same bottleneck afflicts digital pathology and digital cinema—any domain where JPEG 2000 is standard and data volumes have grown faster than the tooling built around them. For CYRAN’s customers in Industry 4,0, in defense and sovereign intelligence, where cloud round-trips are not an option, the only path forward was a local hardware solution capable of decoding gigapixel imagery at interactive speeds. What it needed was hardware that could keep pace with the sensors.
Cyran AI Solutions
Edge AI inference on satellite imagery using Cyran SpatialFuse on NVIDIA DGX Spark.
Eliminating the Hidden Bottleneck With PixelFlux
The answer came in the architecture of NVIDIA DGX Spark. Built on the GB10 platform with an NVIDIA Blackwell GPU and NVIDIA Grace™ CPU sharing a unified LPDDR5X memory pool, DGX Spark offered two properties that no discrete GPU workstation could match for this workload.
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, enabling throughput that CPU-based pipelines simply cannot match at the scale of Earth observation data.
The second was unified memory. On a discrete GPU configuration—even a high-end NVIDIA RTX™ A6000 Ada—decoded image buffers must be copied back to host memory over PCIe after GPU processing. On NVIDIA DGX Spark, the GPU and CPU share the same memory pool, so decoded image data is immediately available without any transfer step. That elimination of overhead grows more valuable as image sizes increase—which is precisely the regime that matters for gigapixel geospatial workloads.
CYRAN operationalized this capability stack in PixelFlux, a GPU-accelerated JPEG 2000 decode library built on NVIDIA nvJPEG2000. PixelFlux replaces the CPU-based decode path entirely, returning decoded image data ready for immediate use—with no transfer overhead on DGX Spark.
For machine learning training pipelines, PixelFlux integrates with NVIDIA DALI to keep the GPU continuously fed from local storage—without CPU preprocessing becoming the bottleneck. The entire solution runs on stock NVIDIA hardware and software, with no third-party infrastructure required.
From Pipeline Stopper to Interactive Analysis
The results are unambiguous. A 26 GB satellite scene that previously required over five minutes of CPU decode time now completes in just over two minutes on NVIDIA DGX Spark—on a system compact enough to deploy at a forward operations center or in an air-gapped, secure environment.
Benchmark Results (3-band uint16 RGB, mean ± std dev, N=10 runs)
“NVIDIA DGX Spark’s unified memory architecture is a revolutionary improvement—removing a fundamental compute bottleneck at the hardware layer from geospatial processing pipelines. Further bundling CYRAN’s indigenized SpatialFuse platform with Spark unleashes new optimized full-stack solutions for end-to-end GeoAI processing, all the way from orbit to ground stations.”
— Dr. Vivek Parmar, CTO, CYRAN AI Solutions
| Image (HxW) | Compressed | Uncompressed | CPU | A6000 Ada | DGX Spark | vs CPU | vs A6000 Ada |
|---|---|---|---|---|---|---|---|
| 2,654 x 6,152 | 38 MB | 93 MB | 1.39s | 0.53s | 0.56s | 2.42x | 1.15x |
| 2,654 x 34,816 | 179 MB | 528 MB | 4.86s | 2.74s | 1.57s | 3.12x | 1.92x |
| 34,816 x 34,816 | 2,862 MB | 6,936 MB | 65.67s | 35.44s | 16.32s | 2.41x | 2.02x |
| 117,504 x 39,168 | 1,333 MB | 26,335 MB | 298.56s | 210.77s | 115.11s | 2.33x | 1.59x |
All times in seconds. DGX Spark leads at every scale.
The performance gap widens with image size—and that is the point. Against the NVIDIA RTX A6000 Ada, NVIDIA DGX Spark wins by 1.15x on the smallest test scene. From a half-gigabyte upward, the advantage grows to over 2x—and it keeps scaling. The larger the image, the more time a conventional workstation spends on data transfer that DGX Spark’s unified memory architecture eliminates entirely.
The business impact extends beyond raw throughput. Development and debugging cycles that previously required cloud infrastructure or multi-GPU servers now run on a desk. CYRAN’s analysts and machine learning engineers can perform genuinely interactive large-scene analysis on local hardware for the first time. For Industry 4.0, defense and sovereign intelligence customers for whom cloud infrastructure is not an option, that capability is not a convenience—it’s a requirement. The results made the case for something larger.
The Edge Extends to Orbit
CYRAN is now building on what PixelFlux demonstrated. The nvJPEG2000 library is portable across NVIDIA Blackwell-based cloud and data center deployments, which means pipelines developed and validated locally on NVIDIA DGX Spark scale to cloud and data center environments without rework—a direct path from edge prototype to production.
Beyond terrestrial intelligence, CYRAN is exploring porting edge AI capabilities to satellite and advanced space payloads for on-orbit inference as per founder Prof. Manan Suri. As the number of space-based sensors expands rapidly, the compute constraints in orbit are even more severe than at the edge—and the same architectural principles that make NVIDIA DGX Spark effective there apply directly: fast local processing, efficient memory use, and hardware built for constrained environments.
As a continuing member of the NVIDIA Inception program, CYRAN is operating at the frontier of geospatial AI, edge computing, and space-based intelligence—and the architecture behind PixelFlux is the foundation it's building on. What began as a five-minute decode problem has become proof that the full stack of geospatial intelligence—from sensor to insight—can run anywhere.
CYRAN AI Solutions is an NVIDIA Inception program member and an IIT Delhi spin-out. Learn more at cyran.in.
Explore how NVIDIA DGX Spark transforms geospatial workflows.