Palette SDK version 1.6

This page lists new features released in SiMa’s Palette SDK version 1.6, including known issues and their workarounds.

New Features and Enhancements

Feature Category	Availability	Feature Description
SDK Installation	GA	Ability to install Docker as part of the installation script (`install.py`).
SDK Installation	GA	Ability to run multiple (upto 4) concurrent GStreamer pipelines.
SDK Installation	GA	Suport for Ethernet-over-PCIe where the default static IP assigned by the module can be overridden by passing `default_ip=false` as a module parameter.
GStreamer Pipeline	GA	Capabilities (CAPS) Negotiation now enabled for data pipelines over PCIe.
SoC Software	GA	Option to update SoC software/firmware when user first connects to the device from the host machine.
Co-processing APIs	GA	Enhanced Python bindings with a Unique ID for each co-processing element, thereby simplifying tracking, referencing, and managing multiple processing elements.
Machine Learning Accelerator (MLA)	GA	Adding a new MLA operator, `Resize` with floating-point scale factors. This allows precise control over input dimensions during ML preprocessing.
Machine Learning Accelerator (MLA)	Beta	Introducing 5D tensor support to the backend, `(N, C, D, H, W)` format, enabling handling of the `D` (Depth) dimension. This release supports medical imaging models like UNet-3D.
Machine Learning Accelerator (MLA)	Beta	Introducing automated surgery of models during import or conversion, thereby transforming and optimizing complex model graphs. This release includes support for the DETR model, extending compatibility to additional transformer-based architectures.

Model Name	Type	FPS
MaskRCNN	PCIe	Achieved 20 FPS performance.
OpenPose	PCIe	Achieved 45 FPS performance.
YOLOv5	PCIe	Achieved 20 FPS performance.
YOLOv7	PCIe	Achieved 120 FPS performance.
YOLOv8 single-Camera	Standalone	Achieved 120 FPS performance.
DETR	Standalone	Achieved 120 FPS performance.
OpenPose	Standalone	Achieved 45 FPS performance.
YOLOv5	Standalone	Achieved 20 FPS performance.
YOLOv7	Standalone	Achieved 120 FPS performance.
YOLOv7 (4-Camera)	Standalone	Each camera runs at 35 FPS, tested with 4 cameras (35*4).
PeopleTracker	Standalone	Achieved 30 FPS performance.
EfficientDet	Standalone	Achieved 20 FPS performance.
PeopleDetector single-Camera	Standalone	Achieved 120 FPS performance.
PeopleDetector (4-Camera)	Standalone	Each camera runs at 35 FPS, tested with 4 cameras (35*4).

Issue	Impact	Workaround
When the system is connected over Ethernet, the network status or interface summary incorrectly displays both Ethernet and PCIe interfaces as active, even if only Ethernet is being used for data transfer.	May cause confusion during system diagnostics or initial setup; tools or logs relying on interface status may misinterpret the actual active link.	Verify the actual data path using system logs or network statistics (ifconfig, ip addr, etc.). Ignore the PCIe interface status if Ethernet is confirmed as the active connection.
In the SDK environment, when a user enters an incorrect password during the `mpk device connect` command, the system does not return an error or timeout. Instead, the command appears stuck or unresponsive until the user manually exits with `CTRL+C`.	During initial setup or device connection attempts users may assume the system is frozen or malfunctioning.	If no response is seen after entering the password, use `CTRL+C` to exit and re-run the command with the correct credentials.