VectorBlox 3.2 General Release Notes
Hardware IP/Reference Design:
- Released the new VectorBlox 3.2 Core IP to Libero catalog, along with an updated CoreVectorBlox IP Handbook for Libero 2026.1.
- Added Unstructured Compression (UCOMP) support for the V500 hardware configuration. Unstructured Compression is now supported on both V1000 and V500.
- V1000 requires a 128-bit AXI width and V500 requires a 64-bit AXI width for both AXI interfaces.
- Reduced resource usage of the V1000 Unstructured Compression core, including approximately:
- 39% lower LSRAM usage (526 → 323)
- 14% reduction in interface logic
For detailed utilization metrics, see resource_utilization.md.
- Updated VectorBlox-SoC-Video-Kit-Demo, VectorBlox-Discovery-Kit-Reference-Design, and VectorBlox-Video-Kit-Demo for VectorBlox 3.2. Please note, the VectorBlox-Discovery-Kit-Reference-Design is only supported for the VBX 3.2 No Compression IP.
SDK/AI Models:
- Replaced
pipwith the uv package manager during environment setup, which speeds up SDK installation significantly. - Added initial support for the transformer models yolo11, yolo26, ViT, Anomaly Detection.
- Added support for TFlite operators BATCH_MATMUL, LAYER_NORM, and improved FULLY_CONNECTED and SOFTMAX performance.
- Added the following new tutorials:
- Transformer tutorials:
- Facial recognition demo tutorials:
- Spacecraft 6D yolo-spn tutorials:
- Added a new yolo-spn demo and updated the facial recognition demo to use the YuNet face detection and SFace face recognition models for the PolarFire SoC Video Kit Demo. For more information, see demo_models.h.
- New C postprocessing types (YUNET, SPACE, and SPACE_T) are documented in C_Postprocessing.md.
- Unstructured Compression compiler improvements:
- Added support for 8x8 and 9x9 kernels (CONV_2D, TRANSPOSE_CONV, and MAX_POOL_2D).
- Increased the sequencer program size to support larger models (up to 2^28 program commands).
- Unstructured Compression models now use the same preprocessing flow (
tflite_preprocess) as the no compression and compression configurations. The--uint8flag is no longer needed in the tutorials. - Replaced the 512×512 Unstructured Compression tutorials with 640×640 cut models:
- yolov8n_50s_25p_cut (replaces yolov8n_50s_25p_512x512)
- yolov9s_70s_15p_cut (replaces yolov9s_70s_15p_512x512)
Unstructured Compression only supports cut models, where post-processing is moved to the host application usingtflite_cut. For guidance and limitations, refer to vectorblox_compression.md and the Unstructured Compression layer limitations in OPS.md.
Bug Fixes:
- Fixed download issues in the kaggle efficientnet-lite tutorials and a cut issue in the onnx/yolov9-t tutorial.
- Added explicit typecasting for NumPy operations to account for overflow.