Xilinx unveiled its Vitis AI software platform back in late 2019, to simplify development for the company’s chip-level hardware architectures and accelerators, using common software development tools and programming languages. Since its unveiling, the Vitis AI development environment has evolved and expanded in scope to incorporate more tools, example designs, libraries and models designed to best leverage Xilinx’s varied array of compute devices, accelerators and Alveo add-in cards. Today, the company is releasing the latest version of the suite, Vitis AI 1.4, and with it comes support for Xilinx’s 7nm Versal ACAP platforms, including the Versal AI Core Series VCK190 Evaluation Kit, the VCK5000 Versal Development Card, and the recently-released 16nm Kria portfolio of adaptive system-on-modules (SOMs).
With this release and the incorporation of support for additional hardware platforms, Xilinx offers somewhat of a one-stop-shop for developers that can address a wide array of machine learning and AI inference workloads, at a multitude of power envelopes and performance levels. The Xilinx Versal AI Core Series VCK190 Evaluation Kit, for example, is capable of MLPerf inference performance levels that far exceed today’s server CPUs. In a ResNet50 closed-division benchmark, the VCK5000 Versal Development Card also outran NVIDIA’s T4 card with a result of 5,921 images per second. And the sub-$200 Kria KV260 Vision AI Starter Kit can be used for vision application development with a, inexpensive, low-power kit that can literally fit in the palm of your hand.
Along with support for additional hardware, a key component of Vitis AI, the AI Model Zoo, has been enhanced. The diversity of the AI Model Zoo has been increased and now features models for 4D radar detection, image-LIDAR sensor fusion, surround view 3D detection, depth estimation, and super resolution.
MORE FOR YOU
Xilinx has made a number of moves over the last few years designed to allow developers and engineers to work in their native AI frameworks, like TensorFlow, PyTorch and Caffe, while targeting the company’s wide array of compute devices and accelerators. Standard programming languages like Python, C++ and OpenCL are also supported and optimized for Xilinx’ hardware. All told, Xilinx’s expanding software ecosystem and breadth of development kits allow developers of virtually all skill levels, to leverage powerful, adaptable accelerators that they may have previously been out of reach. Today’s release of Vitis AI 1.4 further solidifies Xilinx’s position and makes its hardware and software platforms even more accessible to developers from all walks of life.