AMD-Xilinx upgrades Versal ACAP for extreme signal processing - Embedded.com

AMD-Xilinx upgrades Versal ACAP for extreme signal processing

New Versal Premium with AI Engines essentially takes the AI engines block from the Versal AI Core ACAP and drops it into the Versal Premium to provide a combined adaptable hardware platform with a significant increase in signal processing capacity.

AMD-Xilinx has added a new optimized device to its Versal adaptive compute acceleration platform (ACAP) portfolio to address intensive signal processing applications in next generation radar and wireless systems.

Its new Versal Premium with AI Engines essentially takes the AI engines block from the Versal AI Core ACAP (which delivers advanced signal processing) and drops it into the Versal Premium (which features massive DSP (digital signal processing) compute and serial bandwidth capacities), to provide a combined adaptable hardware platform with a significant increase in signal processing capacity.

This allows it to address radar system functions such as adaptive beamforming, signal processing for RF machine learning applications like digital RF memory and direction finding. It also targets a growing demand for wireless testers as part of the global 5G rollout, applications such as 5G protocol testers and production testers, as well as semiconductor automated test equipment. 

With applications in aerospace and defense and test and measurement, the new device is said to deliver a 4X increase in signal processing capacity compared to the last-generation Virtex UltraScale+ VU13P FPGA. It also overcomes I/O bottlenecks with up to 9Tb/s serial bandwidth, and offers significantly reduced size, weight and power through heterogeneous, power-optimized integration of hardened, ASIC-like cores such as 100G/600G Ethernet cores, 400G high-speed crypto engines, DDR memory controller, and integrated PCIe Gen5 blocks.

Xilinx Versal Premium with AI Engines radar application
In radar beamforming applications, the heterogenous compute engines enable 67% smaller footprint and up to 43% lower power with 2X beamforming performance. (Source: AMD-Xilinx)

In an interview with embedded.com, Mike Thompson, the senior product line manager for the adaptive and embedded computing group at AMD-Xilinx, said the new devices address three fundamental challenges facing next generation radar and wireless systems. These are the massive increase in the need for signal processing compute power, limited serial bandwidth preventing higher channel density, and highly constrained size, weight and power. “The addition of AI Engines offers a silicon area efficient additional DSP processing capacity with an array of compute cores that are hardware adaptable to evolving algorithms, with tightly coupled memory.”

Hence by combining its AI Engines module with DSP Engines, users of Versal Premium with AI Engines devices can realize major performance gains compared to prior generation 16nm Xilinx devices and competing products on the market today. The heterogenous integration of both engines enables customers to assign the right compute engine for the right task. In radar beamforming applications, the heterogenous compute engines enable 67% smaller footprint and up to 43% lower power with 2X beamforming performance.

Graphs highlight performance gains for CIN16, FP32, and INT8 acceleration with Versal Premium with AI Engines. (Source: AMD-Xilinx)

Development tools are available for both hardware and software developers. AMD-Xilinx has Vivado ML for hardware developers, and Vitis and Vitis AI development platforms for software developers and AI and data scientists. The Versal Premium series with AI Engines is expected to begin shipping in the first half of next year. However, in order to evaluate it, customers are being encouraged to start prototyping using the existing Versal Premium and Versal AI Core Evaluation Kits and devices for which documentation and tools are available now.


Related Content:

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.