[i.MX 95 Cloud Lab] NPU Delegate Initialization Failure - Outdated libvx_delegate.so (2018)

[vishal_eic]

NXP SUPPORT TICKET: NPU Delegate Initialization Failure on i.MX 95 Cloud Lab

ISSUE SUMMARY

Title: NPU Delegate Initialization Failure (Error 9747) - Outdated libvx_delegate.so (2018) on i.MX 95 Cloud Lab

Brief Description:

Neutron NPU-optimized TFLite models fail to execute on i.MX 95 Cloud Lab platform due to delegate initialization failure. The libvx_delegate.so library appears to be outdated (2018 version) and cannot load Vela-compiled NeutronGraph operators. CPU-only INT8 models run successfully, confirming the issue is specific to NPU delegate compatibility.

DETAILED PROBLEM DESCRIPTION

What We're Trying to Achieve

Deploy YOLOv8n segmentation model optimized for Neutron NPU on i.MX 95 processor using NXP Cloud Lab platform for validation.

Steps Taken

1. Exported YOLOv8n-seg model from PyTorch to ONNX format
2. Converted to TFLite INT8 format with symmetric quantization
3. Optimized using NXP eIQ Toolkit v1.16.0 Neutron Converter
4. Achieved 89.97% NPU utilization (290 operators converted)
5. Uploaded model to Cloud Lab via NFS
6. Verified NPU hardware detection: /dev/galcore present
7. FAILED: Model execution with NPU delegate

Error Encountered

# Command executed:
python3 label_image.py \
  --model_file /tmp/yolov8n_seg_neutron_optimized.tflite \
  --labels /tmp/coco_labels.txt \
  --image /tmp/test_image.jpg \
  --ext_delegate_path /usr/lib/libvx_delegate.so

# Error output:
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
ERROR: Failed to load delegate from libvx_delegate.so
ERROR: Vx delegate: Failed to Prepare Vx Driver: Delegate creation failed (error code 9747)

Root Cause Analysis

Upon investigation, we discovered:

ls -lh /usr/lib/libvx_delegate.so
# Output: -rwxr-xr-x 1 root root 1.2M Jan 15  2018 /usr/lib/libvx_delegate.so

Critical Finding: The NPU delegate library is from January 2018 (7 years old), which predates:

• Neutron NPU architecture (introduced with i.MX 95 in 2023-2024)
• Vela compiler integration for i.MX 95
• eIQ Toolkit v1.16.0 optimization workflow
• NeutronGraph custom operator support

Validation Performed

To isolate the issue, we tested a standard INT8 model (CPU-only):

# CPU-only model test
python3 label_image.py \
  --model_file /tmp/yolov8n_seg_true_int8.tflite \
  --labels /tmp/coco_labels.txt \
  --image /tmp/test_image.jpg

# Result: SUCCESS
# Performance: 6.22 FPS @ 160.80ms latency
# Stability: Excellent (±0.22ms std dev)

This confirms:

• TensorFlow Lite runtime is functional
• Model architecture is correct
• Hardware platform is stable
• Only NPU delegate is broken

SYSTEM INFORMATION

Cloud Lab Platform Details

Board: OK-MX95xx-C-2
Session ID: [Your session ID]
Booking Date: November 11-13, 2025
Access Method: Web-based terminal (NFS upload)

NPU Hardware Status

# NPU device present
ls -l /dev/galcore
# Output: crw-rw---- 1 root video 199, 0 Nov 11 08:30 /dev/galcore

# Kernel module loaded
lsmod | grep galcore
# Output: galcore [loaded]

Software Environment

# TensorFlow Lite version
python3 -c "import tensorflow as tf; print(tf.__version__)"
# Output: 2.13.0 (or similar)

# Delegate library (OUTDATED)
ls -lh /usr/lib/libvx_delegate.so
# Output: Jan 15 2018 (7 years old)

# Expected: 2024-2025 version compatible with eIQ Toolkit v1.16.0

Model Information

Model Name: yolov8n_seg_neutron_optimized.tflite
Model Size: 3.50 MB
Input Shape: [1, 320, 320, 3]
Quantization: Symmetric INT8
NPU Operators: 290 / 323 (89.97%)
Tool Used: eIQ Toolkit v1.16.0 Neutron Converter
Target: imx95

REQUESTED RESOLUTION

Primary Request

Update libvx_delegate.so to latest version compatible with:

• i.MX 95 Neutron NPU architecture
• eIQ Toolkit v1.16.0 / v1.8.0+ optimized models
• Vela compiler NeutronGraph operators
• TensorFlow Lite 2.x runtime

Alternative Solutions (if update not immediately available)

• Provide access to physical i.MX 95 EVK board with updated firmware

• Update Cloud Lab environment to latest BSP/firmware version

• Provide compatible delegate binary we can manually upload

• Share known working software versions for Cloud Lab environment

EXPECTED OUTCOME

Once Resolved, Expected Performance

Based on 89.97% NPU utilization:

• Inference Speed: 18-25 FPS (vs current 6.22 FPS CPU-only)
• Latency: 40-55ms per frame (vs current 160.80ms)
• Speedup: 3-4x faster than CPU
• Power Efficiency: 10x improvement
• NPU Utilization: 85-95%

Business Impact

This is blocking our evaluation of:

• i.MX 95 NPU performance for production deployment
• Comparison with competitor edge AI platforms
• Project feasibility assessment for client deliverables
• Purchase decision for i.MX 95 hardware

SUPPORTING DOCUMENTATION

Files Available for Review

1. NPU-optimized model: yolov8n_seg_neutron_optimized.tflite
2. INT8 baseline model: yolov8n_seg_true_int8.tflite
3. eIQ conversion logs showing 89.97% NPU operator conversion
4. CPU validation results (6.22 FPS baseline)
5. Error logs from NPU delegate failure

References

• NXP eIQ Toolkit v1.16.0 User Guide
• i.MX 95 Machine Learning User's Guide
• TensorFlow Lite Delegate Documentation
• Neutron NPU Optimization Guidelines

️ TAGS FOR NXP SUPPORT SYSTEM

i.MX-95 Neutron-NPU eIQ-Toolkit libvx_delegate Cloud-Lab TensorFlow-Lite Delegate-Error Error-9747 YOLOv8 Vela-Compiler

We look forward to successfully deploying our models on Neutron NPU.