Why Innodisk Chose the Qualcomm Dragonwing™ IQ-9075 Processor for Multi-Channel Video Inferencing and Edge AI

Executive Summary

Innodisk, a leading global AI solution provider, selected the Dragonwing IQ-9075 processor as the foundation for its next-generation Edge AI product family. After extensive benchmarking, Innodisk found that the Dragonwing IQ9 heterogeneous compute engine, combining a dedicated NPU, GPU, and CPU, delivered superior inference throughput in multi-channel, multi-model workloads while consuming less than 30 watts.

The result is a fanless, industrial-grade edge AI platform that is designed to handle 9 simultaneous 1080p video streams running YOLO v10 object detection at 29 FPS average, more than double the throughput of select competing solutions in the same power envelope. For customers across smart factories, smart cities, and autonomous mobile robotics, this can help enable consolidated deployments, potentially reducing system complexity and total cost of ownership, and supporting faster time‑to‑deployment.

The Challenge: Scaling Video AI Without Increasing Power and System Complexity

Multi-channel AI systems create load across the entire pipeline: video ingest, decode, resize, color conversion, inference, post-processing, and display or network output. Adding more cameras increases memory traffic, CPU scheduling pressure, accelerator utilization, and thermal load. For industrial systems, this must be solved inside fixed power and enclosure limits.

Technical requirements

• Sustained multi-stream FPS: Maintain useful frame rates as channel count scales, including 1080p at 29 FPS for 9 channels and 720p at 5 FPS for 32 channels.
• Local inference: Run AI on-device for systems that cannot depend on cloud connectivity.
• Industrial I/O: LAN, serial, CAN bus, and CAN FD for factory and vehicle integration.
• Environmental resilience: Supports wide temperature operation and fanless system designs for harsh environments.

Why Competing Platforms Fell Short

Innodisk's customers evaluating select GPU-based AI modules encountered three blockers that prevented EVK demos from reaching production:

• Multi-channel video discrepancy: Headline TOPS did not translate to concurrent stream performance. Video decode and inference contended for the same GPU, causing frame rates to drop sharply as channels increased.
• Thermal instability: Customer enclosures regularly hit 55 to 60°C ambient. The competing module throttled or destabilized in fanless, sealed industrial chassis and required active cooling that was not deployable.
• Support gap: Engineering issues were routed through distributors and community forums. Customers were told the EVK passed validation and the issue must be in their carrier or peripheral integration, leaving them to debug independently.

The Solution: Dragonwing IQ9 as the Compute Foundation

Innodisk built its EXMP-Q911 and fanless Edge AI platforms around the Dragonwing IQ-9075 and Dragonwing IQ-8275 processors. The Dragonwing IQ-9075 was selected for demanding multi-channel workloads at 100 dense TOPS, while the Dragonwing IQ-8275 offers a 40 dense TOPS option for cost-sensitive deployments. Both platforms use a common COM-HPC Mini module architecture, allowing customers to start with a development kit and move to a deployable fanless edge system without re-engineering their software stack.

Innodisk product options

EXMP-Q911 COM-HPC Mini Production compute module for OEM embedded designs.

EXEC-Q911 Starter Kit. COM-HPC Mini plus 3.5" carrier and cooler for evaluation, software bring-up, and proof of concept.

APEX-A100 Fanless Edge AI Box. Deployable edge AI system (-40°C to 70°C) built for customers that require a complete box-level solution design.

Camera, I/O, Storage and Memory Ecosystem

Innodisk supports the Dragonwing IQ9 platform with a broad ecosystem of camera, connectivity, storage and memory expansion options. This matters for multi-channel AI because this flexibility enables developers to tailor AI vision systems based on deployment requirements, including cable length, enclosure design, synchronization needs, and target environments.

• MIPI CSI-2 camera modules: 2 MP, 8 MP, and 13 MP options with HDR and external trigger support on selected models.
• GMSL camera modules: 13 MP fixed focus, 3 MP fixed focus with HDR LFM, and 2 MP fixed focus with HDR options.
• Industrial I/O modules: LAN connectivity supporting 10GbE, serial communication, CAN Bus, and CAN FD support.
• Storage and memory options: NVMe SSDs and DRAM modules for scalable system performance and capacity.

Software enablement

Two complementary tools cover the model-to-deployment workflow:

• Qualcomm® AI Hub. Model import, conversion, quantization, performance validation, and a library of 250+ optimized open-source models for Qualcomm® hardware. Supported models include object detection solutions such as YOLOv10-Detection.
• Innodisk IQ-Studio (open source).  Quick-start guides, iQS-Streampipe for multi-stream video processing with full GPU/NPU acceleration, iQS-VLM for vision-language evaluation, benchmark and stress testing tools, model conversion documentation, including FP32-to-INT8 quantization and PyTorch-to-TFLite conversion workflows, and peripheral driver integration. Available at github.com/InnoIPA/iQ-Studio.

Together, Qualcomm AI Hub handles model-level optimization while IQ-Studio handles system-level integration: peripheral drivers, application pipelines, and deployment tooling.

Benefit of Dragonwing IQ8 and Dragonwing IQ9 Heterogeneous System Architecture for Inference Pipelines

Dragonwing IQ8 and Dragonwing IQ9 processors assign each pipeline stage to its own dedicated hardware engine. This is the core advantage of the heterogeneous architecture: video decode does not compete with AI inference, and display compositing never stalls pre-processing.

Each pipeline stage runs on a dedicated Qualcomm hardware engine. Hardware acceleration engines operate in parallel, not sequentially on a shared GPU.

Performance Benchmarking

Different vendors calculate TOPS differently: sparse versus dense, floating-point versus integer. Innodisk focused on what production deployments require: end-to-end frames per second measured from video decode through inference to display output.

Multi-channel streaming benchmark

Test: 1080p @ 30 FPS input per channel. YOLO v10n (INT8, 640x640). Full pipeline: H.264 decode, pre-processing, NPU/GPU inference, post-processing, HDMI compositor. Both platforms under 30W.

Single-model inference benchmarks

Test: Both platforms at sub-30W. YOLO series 640x640, UNet-Seg 640x1280, ResNet-50 224x224.

Target Applications

Over 20 customers across Taiwan, Europe, and the United States are evaluating the platform. Examples include:

• Smart factory: SOP inspection and AI NVR: A compact edge system processes camera feeds at or near each workstation for assembly verification, SOP checks, and error detection. Multi-channel FHD analytics run on a single device, reducing response latency and keeping inference local to the production environment.
• AGV and AMR perception: Dragonwing IQ-9075 paired with GMSL cameras provides local obstacle detection, path planning, and navigation for mobile robots. The fanless system option is designed to fit strict vibration, space, power, and thermal constraints without depending on cloud inference.
• VMS and traffic analytics in smart cities: VMS nodes use the platform for scalable multi-camera analytics, from 9-channel 1080p at 29 FPS down to 32-channel 720p at approximately 5 FPS per stream. Roadside and pole-mounted enclosures benefit from the elimination of active cooling as a failure mode.
• Industrial automation and transportation: Flexible I/O expansion (LAN, serial, CAN bus, CAN FD, storage, camera modules) lets teams configure the same Dragonwing IQ9 compute base for different interface requirements across long deployment windows.

Conclusion

Innodisk chose Dragonwing processors because their customers needed efficient AI throughput, environmental resistance, and lower total cost. The heterogeneous architecture in Dragonwing IQ-9075 delivers on all three: 123% higher multi-stream throughput than the tested competing GPU platform, fanless operation from -40 to 70°C, and silicon availability through 2038.

Combined with Innodisk's peripheral ecosystem, direct engineering support, and open-source IQ-Studio toolkit, it is designed to offer the shortest path from evaluation to production for multi-channel edge AI.

Learn More

• Dragonwing IQ-9075 COM-HPC Mini Module: innodisk.com/en/products/computing/qualcomm-solution/exmp-q911
• Dragonwing IQ-9075 Starter Kit: innodisk.com/en/products/computing/qualcomm-solution/exec-q911
• APEX-A100 Fanless Edge AI Box: innodisk.com/en/edge-ai-systems/qualcomm-solution/apex-a100
• IQ-Studio (Open Source): github.com/InnoIPA/iQ-Studio
• Qualcomm AI Hub: aihub.qualcomm.com