ASCEND GPU Pipeline

GPU-Accelerated Visual-Inertial Perception Pipeline for GPS-Denied UAV Navigation

Event: NERSC Open Hackathon, July 15, 2026
Goal: Reduce 65ms CPU pipeline to <11ms on Perlmutter A100

---

Team Structure

Domain	Folder	Owner	Responsibility
A	domain-a-visual-slam/	Paresh	cuVSLAM integration, VIO, trajectory accuracy
B	domain-b-gpu-optimization/	Vishnu	Profiling, roofline, memory optimization, CUDA streams
C	domain-c-ai-inference/	Yugawathi	TensorRT INT8, obstacle detection, model optimization
D	domain-d-robotics-planning/	Kamalesh	Pipeline integration, path planning, SLURM, Perlmutter

Final Target (SCRUM Table)

Stage	CPU Baseline	GPU Target	Speedup
Feature Extraction	28 ms	6 ms	4.7×
VIO / EKF	12 ms	5 ms	2.4×
Depth Estimation	18 ms	3 ms	6×
Obstacle Detection	42 ms	2 ms	21×
Path Planning	8 ms	1.5 ms	5.3×
Transfer Overhead	~5 ms	0.5 ms	10×
End-to-End	~65 ms	~11 ms	~5.9×

Current Status

• Domain B: Phase 1-4 complete (profiling, transfers, streams, roofline, ncu analysis)
• Domain A: Phase 1-4 complete (baseline 1.78ms, 22.5x CPU speedup, cuVSLAM integration plan, roofline, pipeline handoff)
• Domain C: TensorRT INT8 conversion
• Domain D: Pipeline integration + Perlmutter setup

Environment

• Python: 3.11.9 (pyenv)
• Package manager: uv (no conda, no system pip)
• CUDA: 12.8 (PyTorch) / 13.2 (nvcc)
• Target GPU: NVIDIA A100 (Perlmutter), Dev: RTX 4060 Laptop

Quick Start

git clone <repo-url>
cd ascend-gpu-pipeline/domain-b-gpu-optimization
uv sync
uv run python benchmarks/pipeline/pipeline_benchmark.py --frames 100 --save