Xilinx unveils Versal AI Edge SoCs

Xilinx unveils Versal AI Edge SoCs

Jun 14, 2021 — by Eric Brown

— views

Xilinx announced a 7nm, headless “Versal AI Edge” SoC that runs Linux on dual 1.76GHz Cortex-A7 cores and has dual 750MHz Cortex-R5F cores, flexible, FPGA-like “Adaptible Engines” up to 520K LUTS, and “AI Engine-ML” cores up to 479 TOPS.

Xilinx, which is soon to be acquired by AMD, announced its 7nm fabricated Versal ACAP processor family in Oct. 2018 and a year later followed up with more details on the platform’s AI support. Since then, the data center oriented Versal Prime and Versal AI Core series have reached production and the Versal Premium has begun sampling. Now Xilinx has followed up with its latest Versal ACAP. The Versal AI Edge series is the direct heir to the earlier Zynq UltraScale+ MPSoC for Linux-driven embedded computers.

Due to enter production in early 2022, the 7nm Versal AI Edge series spans seven different models from the VE2002 to the VE2802. As with the UltraScale+, the “Scalar Engine” block of Arm Cortex-A and Cortex-R cores is the same on all the models, in this case featuring 2x up to 1.76GHz Cortex-A72 cores that run Linux and 2x up to 750MHz Cortex-R5F real-time cores for RTOS control. (By contrast, the UltraScale+ has 4x 1.2GHz -A53 and 2x 600MHz Cortex-R5 cores.) The main difference here is the lack of an Arm Mali GPU.

Depending on the model, the Versal AI Edge SoCs offer different levels of programmable logic cores, which include FPGA-like Adaptive Engine cores designed for sensor integration. There are also widely varying numbers of “Intelligent Engines,” a tiled array of VLIW and SIMD processing elements that include DSP blocks and NPU-like AI Engine and AI Engine-ML cores.

Versal AI Edge applications include automated driving “with the highest levels of functional safety,” collaborative robots and drones, predictive factory and healthcare systems, and multi-mission payloads for the aerospace and defense markets, says Xilinx. The platform supports advanced signal processing workloads for vision, radar, LiDAR, and software defined radio, and enables applications including machine vision cameras, medical equipment such as ultrasound machines, and various industrial applications.

Versal AI Edge simplified (left) and detailed block diagrams
(click images to enlarge)

The Versal AI Edge SoCs offers 10X greater compute density compared to the 16nm Zynq UltraScale+ and up to 4x the AI performance-per-watt vs. Nvidia’s Jetson AGX Xavier module, claims Xilinx. (The AGX Xavier features 8x 2.26GHz ARMv8.2 cores and a 512-core, 1.37GHz Nvidia Volta GPU with 64 tensor cores that support up to 22-TOPS INT8 performance.)

Total AI performance for INT4 ranges from 14 TOPS to 479 TOPS and for INT8, 7 TOPS to 228 TOPS, claims Xilinx. The AI Engine-ML cores deliver 4X the machine learning compute “compared to the previous AI Engine architecture.” This appears to refer to the Xilinx AI Platform and Xilinx Vitis AI Stack available on the UltraScale+.

The Xilinx AI Platform is a partial backport of the AI Engine-ML available on Versal ACAP (adaptive compute acceleration platform), both of which are based on low-latency AI tech Xilinx acquired when it bought DeepPhi Technology in 2018. The Deephi “sparse neural network” Core technology features a CNN pruning technology and deep compression algorithm to reduce the size of AI algorithms for edge applications.

Major AI, FPGA, and memory specs across 7x Versal AI Edge models
(click image to enlarge)

The Adaptable Engines are a more flexible version of the programmable FPGA logic that sets Xilinx SoCs apart from most other Arm chips. The cores enable from 44K to 1,129K logic cells and 20K to 512K LUTs, depending on the model

Although this is roughly equivalent to the FPGA range found on UltraScale+ EG models, there is more going on here than the specs indicate. The Adaptable Engines implement and adapt sensor fusion algorithms and accelerate pre- and post-data processing across the pipeline. They also enable deterministic networking and motor control for real-time response and isolate safety-critical functions for fail-safe operation. Support is provided for hardware redundancies and fault resilience.

Versal AI Edge specs
(click images to enlarge)

Versal AI Edge devices support multiple safety standards across industrial (IEC 61508), avionics (DO-254/178), and automotive (ISO 26262) markets, enabling vendors to meet ASIL C random hardware integrity and ASIL D systematic integrity levels. These functional safety features are primarily implemented via the Cortex-R5F cores.

Along with the Cortex-A72 and -R5F cores, other Scalar Engine functions include a platform management controller (PMC), which is based on a “triple-redundant processor.” The PMC manages device operation, including platform boot, advanced power and thermal management, security, safety, and reliability. Security features include hardware root of trust with crypto engines, attack protection, and run-time security.

The Scalar Engine block also includes 256KB of on-chip memory (OCM) cache with ECC support. In addition, the four lower-end models feature 4MB (32Mb) of accelerator RAM, which is “accessible to all compute engines and helps eliminate the need to go to external memory for critical compute functions such as AI inference,” says Xilinx. The accelerator RAM can also handle overflow of safety-critical code from the OCM.

One reason the three higher end models lack the accelerator RAM is that they support far more external DDR4 or LPDDR-4266 RAM up to 263Mb (about 32MB). These same models also the only ones to offer CCIX and PCIe with DMA (CPM) and the only ones to provide 4x PCIe Gen4 lanes.

The three high-end models also support far more NoC ports and feature 2x 40GbE multirate Ethernet MACs compared to one or none for the others. They also supply 32 to 44 transceivers at up to 32.75Gbps compared to 8x or none on the lower-end parts. Other peripherals include native MIPI-CSI support for up to 8-megapixel vision sensors.

Versal AI Edge software Tools include Linux and RTOS BSPs plus Vivado design tools for hardware developers. Software developers also have access to the Vitis unified platform, Vitis AI, frameworks such as TensorFlow or PyTorch, and various acceleration libraries.

Further information

Versal AI Edge series design documentation and support is now available to early access customers. Shipments are expected during the first half of 2022. Future models will include automotive- and defense-grade specific models. More information may be found in Xilinx’s announcement and product page.