FPGA IP accelerates neural network computing for edge and embedded AI applications
EP&T MagazineElectronics Semiconductors AI FPGA GPUs IP semiconductor
Global Technology partners Xilinx, Avnet, Libertron and E-Elements to speed development of AI solutions across a range of industries and use cases
AI software innovator Mipsology announced the availability of Zebra FPGA IP, a solution that accelerates the development and deployment of FPGA and adaptive SoC-based machine learning systems. Zebra FPGA IP is optimized to power edge applications spanning industrial automation, security, autonomous vehicles, smart cities, super resolution and more. Zebra IP is currently being used by early access partners and is expected to ship early Q3.
“Mipsology’s solutions continue to lead the path for AI on FPGAs,” said Ramine Roane, vice president of AI Software, Xilinx. “When combined with ARM CPUs and GPU, as well as video decoding, the new Zebra FPGA IP product from Mipsology enables an AI platform, expanding our ability to deliver a superior AI inference solution for our edge and embedded customers.”
Enabling designers to integrate their own functions
The Zebra neural network accelerator can be combined with ARM CPUs and GPUs, and video decoders on Xilinx, Inc. multiprocessor systems on chips (MPSoCs) to deliver complex processing unmatched by conventional SoCs. Zebra FPGA IP uses only a part of the FPGA, offering the highest possible density of computation and enabling designers to integrate their own functions and fully use ARM CPUs/GPUs, delivering high performance and low latency solutions.
“Customers in our region and across the world continue to express a need for superior AI acceleration for edge and embedded AI applications,” said Arthur Chung, Senior Director, Supplier Management, Avnet Asia. “With the new Zebra IP solution from Mipsology, a leader in AI acceleration, we will be able to confidently satisfy customer demand in this area.”
Process many kinds of neural networks
Zebra FPGA IP is based on the proven core technology of Mipsology’s flagship Zebra software. It takes TensorFlow, PyTorch or ONNX CNN models and maps them seamlessly to the target system, while not sacrificing accuracy for the edge. As Zebra can process many kinds of neural networks, the application can continue to evolve without the need to reprogram the FPGA at a low level.
“We are thrilled with the capabilities of Mipsology’s new Zebra IP product,” said MB Kim, CEO, Libertron, Co., LTD. “We can now deliver the best AI inference acceleration solution to our customers who are looking for superior acceleration for their edge and embedded applications.”
Embedded computing market
Zebra IP helps partners expedite the development of neural network-based solutions for the embedded computing market. The growing needs at the edge are not met by available GPUs, which are inflexible and cannot incorporate functions that were not planned when they were designed. The continuous evolution of neural networks also renders ASICs obsolete before they hit the market. The combination of Zebra FPGA IP and the ability to program FPGAs at the hardware level provides an optimal malleable platform for machine learning-based systems at the edge.
“We are very enthusiastic about Mipsology’s new Zebra IP capabilities. Zebra IP will deliver business critical capabilities for AI inference for our Edge and Embedded customers” Said Alex Su, Director, E-Elements Technology Co., Ltd.
Much more flexible format
Zebra FPGA IP is well positioned to power smart systems that use cameras to collect information. When paired with video decoders and ARM CPUs on an FPGA, Zebra FPGA IP can perform complex processing without causing latency. This FPGA solution fits in cameras, set-top boxes, mobile systems, boxes installed in streets, and all other compact formats, without the cooling requirements and limited life spans of GPUs. FPGAs are extremely reliable, requiring minimal maintenance for Edge-based solutions.
“Zebra IP has the same features and behavior as our Zebra inference accelerator running on PCIe cards, but in a much more flexible format, so that developers can enhance their systems with additional capabilities,” said Ludovic Larzul, CEO and founder of Mipsology. “They can integrate complex functions around Zebra to create smart systems deployed everywhere. The possibilities are endless.”