Kneron

K-Scope: Visual Profiler

NeuWare Studio

K-Scope

The MRI scanner for your AI application. Visualize latency, memory, and power consumption with microsecond precision.

User Guide
Supported Devices
KL520 Series
Smart Home / IoT (Always-on)
KL720 Series
High Performance Edge
KL530 Series
Automotive (Transformers)
K-Scope Interface

The Observability Gap

Traditional edge AI debugging involves reading text logs that output "Total Inference Time: 45ms". But this single number hides the truth.

Why is it 45ms? Is it the convolution layers? Is the USB bus choked? Is the host CPU too slow to feed the NPU?

[INFO] NPU Inference Start
[INFO] NPU Inference End (Duration: 45ms)
[WARN] Throughput below target

The Visual Insight

K-Scope instruments the firmware to emit lightweight trace events. It visualizes these events on a unified timeline, similar to NVIDIA Nsight Systems.

  • Unified Timeline: See Host CPU, USB Bus, and NPU Compute on one axis.
  • Concurrency Check: Verify that your data transfer overlaps with compute (Double Buffering).
  • Power Profiling: Correlate high-power spikes with specific model layers.

Targeted Optimizations

Solve specific hardware challenges.

Battery Optimization

KL520 Focus

For battery-operated smart locks, "Always-on" means milliwatts matter.

K-Scope visualizes power draw over time. Identify if your model keeps the NPU awake longer than necessary or if unoptimized layers are causing power spikes.

Bandwidth Analysis

KL530 Focus

Transformers (ViT) are memory hungry.

The DRAM Heatmap shows memory bandwidth saturation. Identify layers that are "Memory Bound" vs "Compute Bound" and optimize channel dimensions accordingly.

System Latency

All Devices

Latency is more than just inference.

Visualize the full "Glass-to-Glass" latency: Camera Capture -> Pre-processing -> USB Transfer -> NPU Inference -> Post-processing. Find the weakest link in the chain.

Automated Bottleneck Detection

You don't need to be a hardware expert to use K-Scope. The "Insight Engine" automatically analyzes your trace and flags common issues.

  • USB Bottleneck DetectedThe NPU is idle for 40% of the time waiting for input data. Consider increasing batch size or using async USB transfer.
  • Thermal Throttling RiskSustained peak power exceeds 2W for >500ms. Consider reducing clock speed or optimizing 'Conv_4' layer.
Interactive Demo

Audit a ResNet50 Trace

We've captured a trace of an unoptimized model. Can you find the memory bottleneck?

Launch K-Scope Mockup