SAN JOSE, Calif.—Xilinx is pushing to expand its presence in the data center with a new line of PCIe accelerator cards powered by the company’s UltraScale+ programmable silicon and aimed at such modern workloads as machine learning inference, video processing and data analytics.
President and CEO Victor Peng announced the Alveo portfolio Oct. 2 during his keynote address at the company’s second Xilinx Developer Forum (XDF) here, noting that it was the first time Xilinx was offering the acceleration capabilities of its field-programmable gate arrays (FPGAs) in a board form factor.
“This gives you the power of adaptability and acceleration without all the work and effort,” Peng said.
In addition, Xilinx is partnering with Advanced Micro Devices to bring together AMD’s Epyc server CPUs built on the company’s “Zen” microarchitecture and Alveo accelerator cards in a system aimed at machine learning inference workloads. The system (pictured) will use two 32-core Epyc 7551 chips and eight Alveo U250 accelerator cards and will be powered by Xilinx’s ML Suite, which supports such machine learning frameworks as TensorFlow. According to officials with both companies, the system has reached an inference throughput of 30,000 images per second on the GoogLeNet convolutional neural network.
The system will help enterprises address the increasing need for more compute capacity at a time when massive amounts of data are being generated and need to be collected, stored and analyzed through such new applications as artificial intelligence (AI) and data analytics.
“New workloads are not just about the CPU; they’re about the system,” Mark Papermaster, CTO and senior vice president of technology and engineering, said after joining Peng onstage at the Xilinx event.
A key message at the show has been the ongoing transformation of Xilinx from a FPGA maker to a platform vendor whose products can help businesses negotiate the rapid changes in the industry with the rise of big data, AI and machine learning, mobility and the internet of things (IoT). Unlike CPUs and GPUs from Nvidia and AMD, FPGAs can be programmed through software, making them much more adaptable to the fast-moving innovation in today’s tech world where workloads can change, standards are introduced and algorithms are updated, Peng and other company officials said.
However, the CEO argued that Xilinx is now more than an FPGA maker, though that message of adaptability continues to be a cornerstone of the company’s portfolio. Over the past few years, the company has also created a family of systems-on-a-chip (SoCs) under the Zynq brand umbrella. In addition, at the show Xilinx introduced its Versal adaptive compute acceleration platform (ACAP) that pulls together multiple compute acceleration technologies, interconnect, memory, management software, software development tools, AI frameworks and other elements into a highly integrated heterogeneous acceleration platform.
The concept of the 7-nanometer acceleration platform was introduced earlier this year in a project dubbed Everest and, like the Alveo accelerator cards, enables Xilinx to better compete with the likes of Intel and Nvidia in the fast-growing AI and big data arenas. Intel bought FPGA vendor Altera for $16.7 billion in 2015.
Organizations over the past 10 years or more have increasingly turned to accelerator technologies, including GPUs and Intel’s Xeon Phi many-core co-processors, to improve the performance of systems without adding greatly to the power consumption. In recent years, FPGAs have been added to the list of chips used as accelerators. Now Xilinx is offering those acceleration and adaptability features in the Alveo cards.
Real-Time Inference Gets a Boost
The company introduced two cards, the U200 and U250. In machine learning, the U250 can increase real-time inference throughput by up to 20 times over high-end CPUs and more than four times for low-latency applications running on GPUs. The Alveo cards also reduce latency threefold over GPUs, which is key when running real-time inference applications, officials said.
Other applications like database searches can run as much as 90 times faster with the accelerator cards than with CPUs.
Inference essentially is one of two parts—the other being training—that make up artificial intelligence workloads. Training involves running massive amounts of data through neural networks to enable them to learn. Inference takes what was learned and puts it to use in systems. It’s similar to students going through years of school to learn and then taking what they’ve learned as they head out on their own into the real world.
The two cards are available now starting at $8,995. Companies also can try them out on the Nimbix cloud.
Peng said there has been considerable interest in the Alveo cards, including from 14 ecosystem partners like CTAccel, Falcon Computing, NGCoded, Xelera Technologies and Bigstream. In addition, Xilinx is working with OEMs including Dell EMC, Fujitsu, Hewlett Packard Enterprise and IBM to qualify the cards on server SKUs.
The work with AMD is an expansion of the relationship between the two companies, according to Xilinx officials. Both have optimized drivers and tuned the performance for interoperability between the Epyc CPUs and Xilinx FPGAs. In addition, they both are working with others on the development of the CCIX interconnect fabric to improve the communication between CPUs and accelerators from disparate vendors.