SC26 · Poster · Accepted

Optimizing HPC Performance for CDC‑Based Storage and Reconstruction of High‑Resolution Imagery

Jooho Kim1  ·  Yifan Yang2  ·  Anish Shakya3  ·  Jacob Kelly4

1Institute for a Disaster Resilient Texas   2Department of Geography   3Department of Marine and Coastal Environmental Science   4Department of Computer Science and Engineering
Texas A&M University

Paper Poster Code
Visual reconstruction check comparing 256 by 256 and 512 by 512 tiled CDC reconstructions of a Galveston drone scene against the original, with absolute difference maps.
Visual reconstruction check. A representative Galveston drone scene reconstructed with 256×256 and 512×512 tiling: original image, CDC reconstruction, and absolute reconstruction-error map. Both tile sizes preserve the fine details—roofs, vehicles, road markings—that emergency managers inspect after a disaster.

Abstract

High-resolution imagery is used more often for community asset management, environmental monitoring, and disaster response. These images preserve fine details such as buildings, roads, vehicles, vegetation, and damage patterns, but they also create a storage problem. A single field mission can generate hundreds of gigabytes of data, which increases long-term storage and transfer costs. To reduce this storage burden, this study evaluates an existing conditional diffusion compression (CDC) workflow for high-resolution image storage and reconstruction on GPU supercomputers. We do not propose a new compression model. Instead, we study how CDC behaves as an HPC workload and identify practical operating points that balance compression ratio, GPU memory use, reconstruction speed, image quality, and downstream analysis utility.

Using 100 RGB drone images from Galveston, Texas, we test different CDC checkpoints and tiled reconstruction settings on NCSA DeltaAI GH200. The balanced checkpoint with 256×256 tiled reconstruction reduces peak GPU memory from 51.96 GB to 1.57 GB and lowers reconstruction time from 143.17 to 79.34 seconds per image compared with full-image reconstruction. This setting keeps a 79.98× compression ratio while preserving useful image quality and downstream detection and segmentation performance. The results show that CDC can support long-term storage of high-resolution imagery when reconstruction is treated as a system optimization problem. A practical HPC setting should report on storage savings, GPU memory, throughput, image quality, visible artifacts, and downstream utility together.

Key Results

The recommended operating point—balanced checkpoint b00064 with 256×256 tiled reconstruction—converts a memory-heavy diffusion job into a low-memory, higher-throughput workflow while preserving disaster-relevant image utility:

52 GB → 1.6 GB
Peak GPU memory
143 s → 79 s
Time per image
79.98×
Compression ratio
0.8705
Detection F1
Selected full-resolution DeltaAI operating points on 50 drone images (Table I of the paper).
SettingSec./imgImg/hrPeak GBRatioF1Roof IoU
High quality 51288.5140.72.9629.73×0.88050.9001
Balanced 25679.3445.41.5779.98×0.87050.8989
Max comp. 25678.9145.61.57139.89×0.86790.8956

Measured Behavior on DeltaAI

Four-panel dashboard of CDC reconstruction performance on DeltaAI GH200: runtime per image, peak GPU memory, compression ratio with fidelity, and downstream detection on human labels.
CDC reconstruction performance dashboard. Measured DeltaAI GH200 results from the N50 compression-setting × tile-size run. Tiling collapses peak GPU memory from 50.74 GB to 1.57 GB and cuts runtime by ~45%, while YOLO detection on human labels stays within ~1.4% of the original imagery.
Heatmap of runtime pressure, GPU memory pressure, non-GPU overhead, quality-loss proxy, detection retention, and SAM roof IoU across four operating points.
Bottleneck analysis from available metrics. Full-image reconstruction is the memory wall (~52 GB peak allocation). Balanced 256 tiling cuts peak memory 32.2× and runtime from 143.17 to 79.34 s/img, while detection retention and SAM roof IoU remain stable across operating points.
Bar charts comparing GH200 and H200 GPUs on fp32 inference time, images per hour, H200 speedup, and peak GPU memory at 65 diffusion steps.
GH200 vs H200 portability check. Matched CDC reconstruction sweeps run unchanged on Delta H200 with a modest ~3.7% speedup over DeltaAI GH200, confirming the workflow ports across GPU platforms; fp16 lowers peak memory from 51.0 to 33.4 GB in the full-image setting.
Table summarizing the evidence scope: DeltaAI GH200 as the main experiment platform, Delta H200 as the portability comparison, and telemetry not yet measured in the repository.
Compute platforms and evidence scope. What each platform contributes: DeltaAI GH200 hosts the main N50 tradeoff, detection, and SAM experiments; Delta H200 provides the portability comparison; hardware telemetry (GPU utilization, power, temperature) is labeled as a future instrumentation step.

Citation

@misc{kim2026cdc_hpc,
  title  = {Optimizing HPC Performance for CDC-Based Storage and
            Reconstruction of High-Resolution Imagery},
  author = {Kim, Jooho and Yang, Yifan and Shakya, Anish and Kelly, Jacob},
  year   = {2026},
  note   = {Poster, SC26: The International Conference for High Performance
            Computing, Networking, Storage, and Analysis, Chicago, IL, USA},
  url    = {https://github.com/rayford295/drone-compression-hpc}
}

Accepted to the SC26 Poster Session (Chicago, IL, November 2026). The formal citation will be updated once the conference entry is available.