Xilinx will today launch a network card that not only offloads acceleration to an FPGA but also to an on-board NXP chip containing 16 Arm CPU cores.
The Cortex-A72 Arm core cluster is expected to run Linux – Ubuntu and Yocto are supported – and act as the control plane, and the gate array as the data plane chugging through packets flowing in and out of the card. The addition of extra CPU oomph is perhaps a sign that customers wanted to run application software and services on the board as well as use its raw hardware acceleration. Red Hat Enterprise Linux and Ubuntu are supported for the host system.
The Alveo SN1000 – dubbed a smart NIC because, well, it tries to add intelligence to network interfaces – is a follow-up to the U25 launched last March and the U50 the year before. The U25 featured a quad-core Arm Cortex-A53 cluster, we note.
For today’s launch, VMware veep Lee Caswell gave a vague-but-positive canned quote about the “composability of Xilinx Alveo SN1000 Smart NICs,” willing us to believe the virtualization giant is still interested in positioning its ESXi-on-Arm hypervisor and guests on smart NICs including Xilinx’s new device, or at least have its software work with the cards.
Meet the ‘DPU’ – accelerated network cards designed to go where CPUs and GPUs can’t be bothered
The aim of this family is allow organizations to fit these cards in servers and similar equipment, and transfer specific workloads to their on-board FPGAs to accelerate said processes in hardware, freeing the host CPU cores of these tasks. What exactly you want to offload is up to you: you could do stateful firewalls, load balancing, IPsec, TLS, NVMe-over-TCP, Virtio.blk storage access, data compression, and so on. This can be programmed in, say, C/C++ and compiled using Xilinx Vitis for the gate array – no need to ruin your day writing Verilog or VHDL, unless you’re into that – or you could drop in software components from Xilinx’s App Store.
The key thing is that this acceleration can be changed on the fly, in a fraction of a second, we’re told, so if the box the card is in is assigned to a new set of work, it can change its programming to adapt and accelerate as needed.
At the moment, these cards are catching the interest of some cloud platforms and hyperscalers, though we’re not sensing too much enthusiasm from enterprises, which do tend to be more conservative in their tech bets. Smart NICs, or data-processing units as some call them, continue popping up from time to time. It’s clear the hardware is there, and software support is peeking around it. Mainstream customer momentum, not so much, we feel. AMD is poised to swallow up Xilinx, and that might be the spark to ignite the FPGA-offload movement.
Speaking of hardware, what’s in the Alveo SN1000? Specifically, Xilinx is kicking off with the SN1022, a full-height, half-length PCIe 4 card due to go on general sale in March. This includes a TSMC-fabbed 16nm UltraScale+ XCU26 FPGA, an NXP processor with 16 64-bit Cortex-A72 cores running at 2GHz with 8MB of cache – nothing to sniff at but certainly not the top of end of Arm’s range – and 4GB of DDR4-2400 RAM for the NXP system-on-chip and 8GB of the same for the FPGA. The XCU26 is said to have a million LUTs.
It has two 100Gbps QSFP28 ports with, we’re told, overall full duplex throughput of 200Gbps, the capacity to process up to 100 million packets per second, TCP throughput of 100Gbps, IPsec encryption throughput of 100Gbps, and a 75W power consumption. There are lots of other bits and pieces, such as hardware-based packet processing, network storage acceleration, virtualization support, a hardware root of trust… a full product brief should appear here sometime today.
But wait, there’s more
Xilinx will also show off a collection of machine-learning software suites that run on its FPGA-based hardware and use the silicon to accelerate tasks that are in vogue right now but give some people the jitters: we’re talking facial and vehicle plate recognition, face mask detection, crowd counting, road traffic monitoring, and object detection and classification. Xilinx spokespeople told us it sells its technology to customers that have legitimate business applications. One of those customers is China’s cloud giant Tencent, which provides the tech to clients to monitor people in buildings via camera feeds and AI.
Below is a table of the hardware you’ll need for each of the suites, which are documented on Xilinx’s smart world webpage.
Table showing Smart World solutions offered by Xilinx and the hardware required for them … Click to enlarge. Source: Xilinx
Finally, the aforementioned Vitis development environment has been updated to include a reference design for building sub-microsecond electronic stock-market trading equipment using Xilinx’s Alveo chips. ®