CHANDLER, Ariz., - Microchip Technology Inc. has released version 3.0 of its VectorBlox accelerator software development kit (SDK), adding support for sparse neural networks to its toolchain for implementing convolutional neural network (CNN) inference on PolarFire FPGAs and system-on-chips.

VectorBlox 3.0 is designed to optimize, compile, and deploy CNN models on PolarFire FPGA and SoC platforms. The SDK works with Microchip's CoreVectorBlox intellectual-property core and is supported by the company's Libero SoC Design Suite.

The new sparse neural network capability allows the accelerator to skip zero-valued operations during inference, reducing the compute and memory requirements of supported models. Microchip says the approach can improve inference efficiency and reduce power consumption for edge AI applications.

The VectorBlox architecture also enables multiple AI workloads, including vision and sensor-processing functions, to run on a single PolarFire device.

Microchip cites space applications as examples of VectorBlox and PolarFire deployments. Planetek Italia used a PolarFire SoC and VectorBlox to process Earth-observation data aboard its AI-eXpress-1 satellite, including object detection and semantic scene analysis. AIKO has also used PolarFire SoC and VectorBlox for its clear_CHARLES software suite, which provides onboard cloud and ship detection for autonomous spacecraft payload operations.

For more information, please visit https://www.microchip.com.