Platform includes updated AI Engine with 4- and 8-bit integer math, along with new memory architecture.
Xilinx has just launched the first edge model of the flexible Versal ACAP (Adaptive Compute Acceleration Platform) family, the third Versal to be announced in addition to the AI, Prime, and Premium models. Versal launched some three years ago and combines flexible reconfigurable logic with hardcoded engines and Arm CPU core. In theory, ACAP’s should be higher performance, more flexible, consume less power, and offer greater ease of programmability for developers than traditional FPGAs. To wit, Xilinx claims the AI Edge can be configured all the way up to 228 TOPS in 8-bit integer performance at 75 watts.
The Versal ACAP is a big deal for Xilinx, and will be for AMD as well. The edge AI market is set to explode, with Gartner projecting some $50B in chipset revenue in 2025. AMD does not enjoy a significant position in the edge, the recent Tesla announcement notwithstanding, so AMD will likely ride Versal into the battle. The ACAP concept is intriguing because it provides a combination of hard (fast) and reconfigurable (flexible) logic, enabling users to tailor an accelerator to specific applications and models. Think of an ACAP as an SoC with FPGA gates for customization. To date, we haven’t seen much traction for Versal, but it is still early. The Cambrian Explosion didn’t happen in a year or two.
What did Xilinx announce?
When Xilinx first announced the ACAP strategy in 2018, the company laid out a roadmap that included six different design points: the AI Edge, AI Core, AI RF, Prime, Premium, and HBM series. Of these markets, Edge has the largest opportunity, in our opinion, so getting this product right with a long lifespan is critically important to Xilinx and thus AMD. It looks it looks to us that the new AI edge platform may just fit the bill, but not until 2022.
Xilinx claims the Versal AI Edge platform quadruples the efficiency of a GPU, and increases compute Xilinx
Xilinx upgraded the AI Engines for faster inference processing at the edge, with enhanced performance from reduced precision operations (four and eight bit integers), and added a memory hierarchy that should enable larger models to run on this platform. The compute density appears exceptional, delivering 10 times density of the previous FPGA implementations from Xilinx. The 2022 availability may seem somewhat disappointing, but the products these platforms will go into have multi year design cycles, so letting customers know what is coming down the road is in both Xilinx and their clients’ best interest.
Xilinx release slides showing how the company envisions developers can deploy the entire application on a Versal AI Edge. Below is the layout for autonomous vehicles showing where workflow elements such as perception (Arm), environmental characterization and sensor fusion (Adaptable Engines), image conditioning (DSP) and object classification (AI engines) could all run simultaneously on the ACAP. The company shared similar maps for factory automation, collaborative robotics and unmanned aerial vehicles. These diagrams make the point that you do not necessarily need other chips for a complex workflow to execute in real-time, as the ACAP Swiss Army Knife has one of each class of domain-specific architectures (DSAs).
Xilinx showed how an AI Edge platform could support a wide variety of computational tasks, iXilinx
Speaking of DSAs, Xilinx also touted the ability to go beyond static optimized engines with what they termed Dynamic Function Exchange, wherein the FPGA-portion of the ACAP can be reconfigured on the fly in milliseconds to do a different job. Imagine how this could be used to reconfigure the device in a vehicle for drive mode with lane management, low-speed mode for parking assist, and post-drive mode for safety and security. As developers begin to deploy this capability, we believe it will give Xilinx and AMD a unique and highly differentiated advantage beyond the usual power, performance, and area dimensions of silicon platforms.
The AI Edge development tools will be available in the 2nd half of 2021, with early silicon shipping in the first half of next year and an evaluation kit and volume production starting in the 2nd half of 2022.
Conclusions
It has taken a long while for the Versal ACAP family to get out of the gate and into production. Based on this announcement, it will still be a while. But those who want to take advantage of the unique features will want to start now so that their end-products can be ready in 2023. While not the fastest nor most power-efficient platforms we have seen, the Versal AI Edge series offers vastly more flexibility and customization potential. Some things are worth the wait.