Xilinx has launched a range of data center products based on ‘composable hardware’ to enable greater adaptability and flexibility to tackle constantly changing data center needs, as well as addressing latency issues in video analytics and financial trading platforms.
The products include a new family of Alveo SmartNICs (network interface cards), artificial intelligence (AI) video analytics applications, an accelerated algorithmic trading reference design for sub-microsecond trading, and the Xilinx App Store. The aim of the new products is to enable software developers to provide compute-intensive performance without requiring expensive and lengthy hardware development. With these new products and solutions, Xilinx said it is eliminating the barriers for software developers to quickly create and deploy software-defined, hardware accelerated applications on Alveo accelerator cards.
The challenge for many data centers is the massively increased demand for networking bandwidth and to optimize for AI and real-time analytics workloads, which are constantly evolving and pushing existing infrastructure to its limits. According to Xilinx, this is driving the need for fully “composable”, software-defined hardware accelerators with the adaptability to optimize for demanding applications as well as the flexibility to quickly take on new workloads and protocols, and accelerate them at line rate.
The first of its new suite of products for the data center is the Xilinx Alveo SN1000, a family of composable SmartNICs. By composable, the company means it offers software-defined hardware acceleration for all function offloads.
In a briefing with embedded.com, Kartik Srinivasan, director of product marketing for the data center group at Xilinx, said, “In our engagement with hyperscalers and early adopters, everyone has different needs, but all need continuous feature innovation, while not having to compromise on performance. With the NIC being a critical component in the I/O of a server, beyond 25Gb, offload NICs started emerging.”
“With ASIC implementations of these offload NICs, there is no programmability and hence no way to scale and adapt for different needs. Similarly, for CPU implementations, while some may have aspects of programmability, they are not efficient and can’t scale to 100Gb connectivity and beyond.”
The SN1000 SmartNICs directly offload CPU intensive tasks to optimize networking performance, with an open architecture that can accelerate a broad range of network functions at line rate. They provide software-defined hardware acceleration for a wide range of networking, security, and storage offloads, such as Open vSwitch and virtualization acceleration (Virtio.net). Security offloads include IPsec, kTLS and SSL/TLS and accelerated storage applications including Virtio.blk, NVMe over TCP, Ceph, and compression and crypto services.
Using the Vitis Networking platform and industry standard, high-level programming languages such as P4, C, and C++, software developers can create network functions, protocols, and applications that operate in hardware on the SmartNIC. Vitis Networking allows organizations to quickly and easily compose new, and tweak existing, network functions to handle new protocols and applications without replacing hardware, future proofing investments.
Srinivasan added, “Customers get an out-of-the-box plug and play experience. We program some features which we think customers require, plus acceleration, and then customer can remove blocks or add their own custom accelerators.”
At launch, Xilinx said it has good industry support for its SmartNICs. Lee Caswell, vice president of marketing at VMware’s cloud platform business unit, said, “VMware is defining the hybrid cloud architecture for next-generation applications with heightened security requirements. SmartNICs will play a critical role in the VMware cloud foundation architecture by giving customers a unified management, security, and resiliency model that spans both bare metal and virtualized environments where the composability of Xilinx Alveo SN1000 SmartNICs will provide a flexible, integrated, and qualified solution for customers.”
Ben Li, general manager of the network business unit at Inspur, added, “Today’s software-defined data center demands flexibility and scalability to meet the ever-changing needs of customer workloads. The innovative composable offload framework of the Xilinx Alveo SN1000 SmartNIC family gives Inspur the agility to adapt to evolving customer needs by rapidly developing and deploying custom workloads.”
Based on the Xilinx 16nm UltraScale+™ architecture, the SN1000 family of SmartNICs are powered by the low-latency Xilinx XCU26 FPGA and a 16-core Arm® processor. SN1000 SmartNICs deliver dual-QSFP ports for 10/25/100Gb/s connectivity with leading small packet performance and a PCIe Gen 4 interconnect. The first model in the family is the SN1022, which is offered in a full height, half-length form factor in a 75-Watt power envelope.
Video analytics platform targets latency-sensitive AI video inferencing
The second of Xilinx data center products to launch is its smart world video analytics platform, along with an ecosystem of partner solutions built to accelerate the most complex and latency-sensitive AI video inferencing applications.
Powered by the Xilinx video machine-learning streaming server, it delivers whole application acceleration and can support multiple neural networks on a single Alveo accelerator card at deterministic sub-100ms pipeline latency. Xilinx said the result is the industry’s lowest TCO (total cost of ownership) for demanding AI video analytics applications.
The partner solutions available at launch include:
- Aupera: offering turnkey smart city and smart retail solutions that combine Aupera’s intelligent, video processing with Alveo accelerators.
- Mipsology: providing a toolset for easy migration of existing AI applications from GPU-based architectures to the Alveo platform, as well as plug-and-play, high-performance AI inference acceleration.
- DeepAI: delivering AI training at the edge on Alveo accelerators with up to a 10x performance per cost advantage compared to GPU-based solutions.
Xilinx told embedded.com that the retail market is a key vertical target for this product platform, given that shrinkage from theft is a big issue in the sector.
Reference design for low latency electronic financial trading
Achieving submicrosecond latency in algorithmic trading has traditionally required costly and time-consuming hardware development. The Vitis development platform now includes an accelerated algorithmic trading (AAT) reference design which gives software developers the ability to quickly and cost-effectively deliver sub-microscond trading performance without the need for custom hardware development.
AAT, implemented on Alveo accelerator cards, provides a modular design that includes all the necessary components for an end-to-end, low-latency trading solution. Each module can be customized using C and C++ in the Vitis platform to meet each firm’s specific needs. The AAT reference design is available today at no cost for Alveo accelerator card customers.
App store with ready-to-deploy video analytics and financial trading solutions
Completing its lineup, Xilinx has also launched an app store with ready-to-deploy accelerated applications spanning smart world AI video analytics to anti-money laundering and live video transcoding. Developed by Xilinx ecosystem partners, the containerized pre-built applications provide an easy way to evaluate, purchase and deploy accelerated applications in minutes.
- Optimizing AI for embedded applications
- Hardware accelerators serve AI applications
- Edge AI platforms beat GPU/CPU on performance and efficiency
- Transforming chip and system communications