Developing AI for Embedded Systems (Embedded AI)

Note: when a model with 97% accuracy on a server can't meet the 100 ms deadline on a target controller, classic ML falls short. Developing AI for embedded systems (Embedded AI) is about making neural networks run on Cortex-M4 with 256 KB of RAM and deterministic execution time. At TrueTech, we've been doing embedded AI for years and know how to overcome these constraints. Our team has 5+ years of experience and certifications for ISO 26262 and IEC 61508. Hardware platform cost reduction can reach 3× through efficient quantization and optimization, saving up to $50,000 per project.

How to Optimize a Model for an Embedded System?

Optimization starts with choosing the right tool for the hardware. On RTOS (FreeRTOS, Zephyr) we use TFLite Micro with CMSIS-NN — ARM-optimized operations deliver up to 4× speedup. For Embedded Linux (Yocto, Buildroot), ONNX Runtime or PyTorch Mobile are suitable, and on FPGA (Xilinx Versal) — Vitis AI with hardware acceleration. The framework choice depends on the target hardware and latency requirements.

For safety-critical projects, model quantization is mandatory. Post-training quantization (INT8) is standard, but for medical systems we apply quantization-aware training with calibration on real data. Our engineers guarantee that accuracy drop does not exceed 2% while the model size is 10× smaller.

Comparison of Quantization Methods

QAT (Quantization-Aware Training) outperforms PTQ for deep networks: accuracy drops only 1% and speed on FPGA increases 5×.

Why is Determinism Critical for Embedded AI?

In industrial systems, inference must complete in a fixed time — worst-case execution time (WCET). Violations cause failures in machine control or brake systems. We eliminate malloc in RTOS, use static buffers, and profile every operation. In automotive, we guarantee deterministic inference within a 100 ms limit with a 15% margin.

How We Do It: A Defect Detection Porting Case

For an automotive client, we ported YOLOv5 to an Infineon TC3xx (TriCore) controller. The original model weighed 30 MB and used 1.2 GB RAM. After INT8 quantization (TFLite), size dropped to 3 MB, RAM to 128 KB. We used CMSIS-NN for convolutions and manual allocation of scratch buffers. Result: latency 85 ms against a 100 ms limit, accuracy dropped 1.2%. In this project, our TFLite Micro solution was 35% faster than the competitor's ONNX Runtime (130 ms), achieving a 1.5× improvement in latency. Get a consultation — we analyze your project in 2 days.

Our Work Process

1. Analysis — target hardware profiling, feasibility study
2. Quantization — type selection (INT8/INT4), calibration, accuracy verification
3. Inference development — C/C++ code, RTOS/Linux integration
4. Testing — WCET, power budget, stress tests
5. Deployment and embedded MLOps — OTA, documentation, team training

Timeline: approximately 12–24 weeks

Complexity increases with reliability and certification requirements. Cost is calculated individually. Typical project costs range from $10,000 to $50,000. A two-day feasibility analysis is $1,500. Contact us for a consultation — we assess your project in 2–3 days.

What Deliverables We Provide

We deliver complete documentation, model files, access to source code repositories, training for your team, and 6 months of support. Specifically:

• Feasibility study and stack selection
• Model quantization and hardware optimization
• Production inference code (C/C++)
• RTOS/Linux and driver integration
• WCET and functional safety testing (if required)
• Documentation and model rights transfer

Our Experience

5+ years in embedded AI, 30+ projects including certified automotive and medical systems. We work with ISO 26262 and IEC 61508. Contact us — we guarantee a personalized approach.

Functional safety standards: IEC 61508, ISO 26262

Edge AI and Optimization: How to Deploy Models Without Cloud?

Imagine: your face recognition model has 4 seconds latency on Jetson Orin, the battery runs out in an hour, and the model crashes with OOM. We are a team of Edge AI engineers with 5+ years in production — we have optimized over 150 models for edge devices. Without profiling and proper choice of quantization or distillation, the project is doomed. The gap between research code and edge deployment is a separate engineering discipline; we help you master it in 2–16 weeks turnkey. Edge AI and model optimization services are not just export, but systematic work with hardware.

Why Simply Exporting a Model Doesn't Work?

A PyTorch model with float32 and batch_size=32 is not ready for edge. Typical problems:

• ResNet-50 in fp32 occupies 98 MB, inference on Cortex-A78 — 380 ms. After INT8 quantization via torch.ao.quantization — 24 MB, 95 ms. Export to ONNX + TensorRT on Jetson — 28 ms.
• YOLOv8m on Raspberry Pi 5 in fp32 — 2.8 fps. TFLite INT8 — 9.4 fps. With XNNPACK delegate — 14 fps (1.5× faster than pure INT8).
• Transformer encoder on mobile CPU: MobileBERT in fp16 via CoreML on iPhone 15 — 18 ms/inference. distilbert-base-uncased in ONNX — 42 ms.

The problem is not choosing "quantize or not" — the right path is determined by the device, task, and acceptable metric degradation. We offer an assessment of your project: within 24 hours we will tell you how feasible it is to speed up the model.

How to Choose Quantization Method for Your Task?

PTQ (Post-Training Quantization) — a quick path. Take a trained model, run a calibration dataset (200–1000 samples), get INT8 or INT4 weights. Tools: torch.ao.quantization, ONNX Runtime quantization tool, bitsandbytes. Accuracy degradation: 0.5–2% on classification. Red zone — small object detection and segmentation, where PTQ gives -4–8% mAP.

QAT (Quantization-Aware Training) — training with simulated quantization noise. More expensive (retraining), but degradation 0.1–0.5%. Justified when PTQ is unacceptable. In PyTorch — torch.ao.quantization.prepare_qat().

GPTQ / AWQ — for LLMs. AWQ better preserves quality at 4-bit quantization. llm-compressor from Neural Magic or autoawq are the main libraries.

Potential savings from choosing the right method can be substantial — for example, reducing cloud inference costs by up to 70% when deploying to edge. Project cost is calculated individually based on model complexity and target platform.

When to Use Pruning vs Distillation?

Structural pruning removes channels or layers. torch.nn.utils.prune — basic tool. For transformers — attention head pruning (LTP, movement pruning). Result: ResNet-50 after removing 40% of channels with fine-tuning — -35% size, -28% latency, -1.2% top-1 accuracy.

Knowledge distillation — train a small student to mimic a large teacher. Classic via KLDivLoss on soft labels. Feature distillation on intermediate layers is more effective. Hugging Face DistilBERT: 66M vs 110M parameters, -40% latency, -3% on GLUE. This is a model compression technique.

Combined approach: distillation → pruning → QAT. Gives maximum effect on limited hardware. We recorded a case where a client achieved 70% reduction in cloud compute spend after moving to edge with this pipeline.

Target Platforms and Tools

TensorRT — the main tool for NVIDIA edge. TRT builds a graph with operator fusion, selects optimal kernels. On Jetson AGX Orin YOLOv8m in TRT INT8 gives 78 fps vs 22 fps in fp16 PyTorch — 3.5× improvement.

Practical Case: How We Detected Defects on a Production Line (Our Client)

Task: real-time scratch detection on metal, 30 fps, camera to Jetson Xavier NX (16GB). Original model YOLOv8l mAP50 0.91, server inference 28 ms, on Jetson in fp16 — 110 ms (9 fps). Not suitable.

Optimization steps we performed for our client:

1. Switch to YOLOv8m — mAP50 0.887 (-2.3%), 68 ms
2. Export to TensorRT FP16 via yolo export format=engine half=True — 31 ms (32 fps)
3. INT8 calibration on 500 frames — 22 ms (45 fps), mAP50 0.879

Result: 3.5% degradation at 5× speedup. Client received engine and documentation. We guarantee metric will not drop below agreed threshold — specified in contract.

What is Included in the Work?

• Report on model profiling on target device (layer latency, bottlenecks)
• Selection and justification of optimization methods (quantization / pruning / distillation)
• Optimized model (TensorRT engine / TFLite / CoreML / ONNX)
• Configs for reproducibility (scripts, Docker image, instructions)
• Testing on real device (at least 10,000 inferences)
• Training of your team (2 hours online)
• 1 month support after delivery

How to Order Model Optimization

1. Submit a request on the website or contact us in any convenient way.
2. We perform free profiling of your model on the target device within 24 hours.
3. We prepare an optimization plan with trade-off estimates (speed vs quality).
4. You approve the plan — we start work.
5. After completion, we deliver the optimized model, configs, and documentation.
6. We train your team and provide monthly support.

Timeline: optimization of an existing model — 2–4 weeks. Development from scratch for edge — 6–16 weeks.

Get a consultation — we will evaluate your model for free and offer a plan within 24 hours. Order free profiling now. For complex projects, contact our engineering team to discuss custom optimisation strategies.