LogicTronix has deployed dozens of custom neural networks using Vitis AI 6.1 and the NPU flow on AMD Versal ASoC platforms. We are specialized in deployment and acceleration on the VEK280 , other Versal boards and wide range of MPSoC devices. Our expertise includes optimizing and deploying CNNs, RNNs, and language models on the Versal platform using the AMD-Xilinx toolchain—such as Vitis AI, NPU flows, as well as our own custom acceleration frameworks.
PointPillars turns point clouds into vertical columns (“pillars”), learns features for each pillar from stacked pillars, generates 2D pseudo image, and applies 2D CNNs on pseudo image to detect 3D objects.PointPillar is one of the largest detection networks and requires significant computational resources to operate in real time. We have successfully deployed PointPillar on the VEK280 platform in a real-time setting by applying multiple customizations and modifications to both the neural network architecture and its inference pipeline.
In the tutorial, we are showcasing PointPillar network Quantization/Compilation (Snapshot creation) for Versal VEK280 with Vitis AI 6.1 , NPU flow.
For complete steps, please follow our Hackster.IO tutorial – https://www.hackster.io/LogicTronix/preparing-pointpillars-network-for-vitis-ai-npu-and-versal
The output of the PointPillar Network from Versal VEK280 board looks like:
ADAS Sensor Fusion Solution with Vitis AI 6.1 NPU and Versal VEK280
In our ADAS sensor fusion demo, we leverage multiple neural networks alongside multiple NPU, AIE, and PL-based processing instances to achieve a high-throughput, high-performance solution.