Xilinx is expanding its support for AI applications from the cloud to the edge with the latest product in its Versal AI adaptive compute acceleration platform (ACAP) portfolio.
The company is launching the Versal AI Edge series with particular focus on striking a balance between high performance per watt, low latency, and the low power and low thermal requirements specific to end point devices in sectors like the automotive and robotics industries.
AI use cases were not extending that far out from network cores and clouds when Xilinx unveiled its first Versal ACAPs in 2018, but an edge product was always on the Versal roadmap, according to Rehan Tahir, senior product line manager at Xilinx. Plus, one of the benefits of ACAP technology is that changes can be made at both the hardware and software levels to better support different kinds of applications and workloads.
“The concept of edge AI is becoming more mainstream now,” Tahir told Fierce Electronics. “And there are some similarities to cloud AI, as well as some differences.” AI in both cloud and edge applications requires high performance, but at the edge there are additional concerns: low latency, low power--which means a need for high performance per watt--and low thermal management (because they may lack the cooling capabilities common in data centers.)
“A lot of edge use cases need massive AI compute, but they are power-constrained and thermally constrained,” Tahir said. “They also have lots of end points requiring sub-five-millisecond latency, which the cloud cannot provide. The edge also has unique safety, security and data privacy issues because in a car, for example, the AI is very close to humans.”
Tahir said Versal AI Edge meets those requirements by leveraging the production-proven 7nm Versal architecture and “miniaturizing” to support AI compute at low latency, power efficiency as low as six watts and with the safety and security measures required in edge applications. At the same time, Versal AI Edge was able to hit the goal of high performance-per-watt, achieving four-times higher performance-per-watt than published benchmarks for Nvidia’s Jetson AGX Xavier GPUs, along with 10-times greater compute density versus previous-generation adaptive system-on-chip products, Tahir said.
The Edge series is rolling out to some early access customers, and will be ready to ship in the first half of 2022. There is also a roadmap to support automotive and defense-grade devices, the company said.
Dan Mandell, senior analyst for IoT and embedded technology, agreed with Tahir's assessment that the time is right to bring better support for AI out to the edge. He told Fierce Electronics via email, "This is a natural and rightly-timed evolution for Xilinx to make with the emergence of edge computing and AI. The edge and intelligent end points are rapidly evolving to support increasingly sophisticated (AI) workloads that demand greater processing performance and hardware/software flexibility. The FPGA fabric features a number of unique advantages over alternative hardware accelerator architectures with regard to programmability and lifecycle support that Xilinx can take advantage of for the edge."
He also said that the Versal AI Edge series could prove even more significant to AMD as that company looks to complete its proposed acquisition of Xilinx. "The Versal family is a critical aspect of that deal and in particular advancing (and diversifying) AMD’s play in embedded and edge AI, which to this point has not been as aggressive as others in the market," Mandell stated. "With Xilinx, AMD gains a significant advantage supporting a much wider range of emerging high-performance embedded and edge (AI) computing opportunities."
RELATED: Xilinx joins the AI chip race