SAN JOSE — When it comes to programming FPGAs, most people think you need to be a chip designer or a step up from being able to program a quantum computer. But Xilinx wants to change all that.
At the Xilinx Developer Forum (XDF), the company announced a completely new development platform called Vitis — and the new platform is going to be open and free.
This is a major change for a company whose primary development platform Vivado, is geared toward low-level control over FPGA configuration. FPGAs have always been more adaptable, but now Xilinx wants them to be more easily programmable. And not just FPGAs like Zynq — the real inspiration for Vitis is Xilinx’s new Versal heterogeneous processor. Versal is a new type of processor, introduced at last year’s XDF, that incorporates CPUs, DSPs, FPGA fabrics, on-chip networks, and machine learning accelerators, making it a performant, versatile, and adaptable processing chip.
The key to successfully use any FPGA, whether it’s in embedded applications or in a cloud services, is the software tools. That is where Vitas is critical for wider acceptance of FPGAs and the Versal architecture.
The Vitis platform is a unified tool that can take Xilinx HLS or C/C++ code and target any processing element from within a standard development environment. Xilinx has committed to support Vitis as free and open. The company built an extensive list of 400+ library functions available on Github. There will also be a Vitis AI tool with support for standard frameworks like TensorFlow, Caffe, and PyTorch. One of the benefits of FPGAs is that you can change the hardware to fit the AI model.
The goal for Vitas and Versal is to move Xilinx beyond being an FPGA vendor, to a flexible and adaptable processing vendor. Xilinx is betting heavily on heterogeneous computing and in order to make that happen, you need a unified software architecture that programmers can use for the various elements inside the Versal chip. The idea that Xilinx would make the software open, and free is a radical re-imaging of how Xilinx goes to market.
Vitis is in beta with some key customers with general availability scheduled for November 1. Check Developer.xilinx.com for more information.
Xilinx Growth Markets
The three major growth drivers for Xilinx are 5G infrastructure, automotive platforms, and data centers. While traditional FPGAs have made good progress in all three areas, Versal is a more complete platform that can run standalone, not just as an accelerator to a host processor. That said, Xilinx recognizes that a couple of Arm cores may not be enough for some applications and Versal is equipped to support PCIe 4.0, and the Vitas environment is designed to support multiple types of host processors including the on-chip Arm, x86, and Power.
The advantage to using a programmable part like Versal is that it allows designs to change more rapidly, even in the field. One of the key application markets is machine learning inference in data centers where algorithms are continuing to change. In addition, even cloud providers are moving toward becoming heterogeneous computing platforms. Amazon’s AWS cloud service has a number of different heterogeneous compute instances, including FPGAs and GPUs. New at XDF, Xilinx announced that it is also added Microsoft’s Azure cloud services — a big win for the company. With the addition of Azure, Xilinx has FPGA instances in the two leading cloud providers.
In the communications market, 5G deployment is a key opportunity for Xilinx — from RF to core routers. For 5G, massive MIMO 16×16 antenna arrays, designed for spectral reuse, require advanced calculations and even machine learning to optimize beam forming. Xilinx has designs with communication companies such as Samsung and Huawei.
Automotive is another market segment Xilinx is focused on. Here there is plenty of opportunities from data aggregation and pre-processing of sensor data, sensor fusion, ADAS to up to autonomous Level 4. In fact, Xilinx already has significant number of design wins in automotive and shipped 60 million units for ADAS designs. Xilinx already has a big business in automotive and claims to be number two in forward camera processing and number one in LiDAR and 4D radar. FPGAs offer design flexibility and low latency processing.
The cutting edge of autonomous driving will probably be “robotaxis” and Xilinx had a very interesting startup on stage called pony.ai that is addressing this market. Pony.ai demonstrated its software’s ability to navigate a car through busy Chinese traffic. Automotive and autonomous vehicles are still evolving, and the innovation cycles are outpacing the silicon design cycles, making it an opportunity for FPGAs and Versal. For more conventional approach to ADAS, Hitachi Automotive Systems is using Xilinx FPGAs for ADAS Level 2+ and has begun to invest in the Versal designs.
While consumer autonomous driving may be a few years off, ride share companies like Uber and DiDi would benefit greatly from robotaxis. Xilinx showed data that estimates an autonomous cab will cost ~$0.60 per mile to operate and be able to run up to 30 rides per day, while a human-driven car costs about $1.50/mi and can make about 15 trips per day. The economics of autonomous robotaxis is compelling for the transportation-as-a-service business.
While some people in the chip industry are pushing the idea of domain-specific processing architectures, those people often overlook the idea of reprogrammable logic such as FPGAs, and Versal. While both are heterogeneous computing architectures, domain-specific designs lead to chips or 3-D package chips with lots of excess silicon that may be used from time to time, but not all the time. In the case of FPGAs, it can be more efficient at re-utilizing silicon by reprogramming it based on the need of the moment. This allows the silicon to be shared in the time domain for different functions. Hybrid architectures utilizing FPGA technology has the potential to be more silicon efficient.
All these new ideas for chip architectures are coming about due to the end of Dennard scaling and the slowing of Moore’s Law. No longer can we rely on process improvements to make chips faster, lower power, and cheaper. New architecture approaches are now needed to bring us the advance processing capabilities we need for new challenging workloads and Xilinx has a unique solution that can scale that will now be easier for more mainstream programmers to use.
Kevin Krewell is Principal Analyst at Tirias Research. Previously, he was a Senior Analyst at The Linley Group and a Senior Editor of Microprocessor Report. He has more than 25 years of industry experience in both engineering and marketing positions including Director of Strategic Marketing at Nvidia and Director of Technical Marketing at Raza Microelectronics (now part of Broadcom). He earned a BS in electrical engineering from Manhattan College and holds an MBA from Adelphi University.
>> Continue reading the complete article originally published on our sister site, EE Times: “Open Software Stars at Xilinx Developer Forum.”