Xilinx is well known as the company that invented the field-programmable gate array (FPGA) in 1985 and then promoted the fabless chip manufacturing model.
In 2020, all these years later, the San Jose-based company wants its name to be equated with ‘adaptability’ as in adaptable hardware architectures and its Adaptive Compute Acceleration Platform (ACAP). It was introduced in the form of the Versal chip in 2019 that is manufactured by TSMC on a 7nm process.
“Our fundamental programming model is to be able to adapt hardware to the software you are writing,” said Emre Onder, Senior Vice President of marketing for Xilinx, in an interview with FierceElectronics. “We can apply the right level of memory and processing to the data…and we can pipeline the processing and run it through a device to give performance with very, very low latency.”
The ACPL-352J 5 A gate drive optocoupler device has integrated fail-safe IGBT and MOSFET diagnostics, protection, and fault reporting. The device features CMTI greater than 50 kV/?s and provides minimal propagation delay with excellent timing skew.
ACAP is Xilinx’s attempt to future-proof semiconductors as Moore’s Law slows down. “We recognized what was happening quite a while ago, and we provide an architecture that doesn’t rely on Moore,” Onder added. Artificial intelligence engines and other elements are hardened in the chip. “If we relied on Moore at 7nm, we wouldn’t have seen the same benefits…ACAP is fairly new, and it’s going to have many years of life.” With a slate of 4,400 patents under its belt and decades of experience, Xilinx arguably understands technology demands and trends.
The value of ACAP to developers is that while there could be 36 billion or more transistors in a chip, most developers don’t want to (or can’t afford to) program at the hardware level. “That’s why there’s software for developers that don’t want to be in the weeds of hardware design,” he said.
That value proposition seems to be resonating with early customers but will be a heavy lift for a company known as the FPGA innovator.
“Versal will take time to build up steam,” said Kevin Krewell, an analyst at Tirias Research. “It’s still early and still rolling out. Versal is definitely different and it’s more System-on-Chip with capabilities for machine learning. Xilinx hired new architects into the company to create more complex systems on chips. Instead of a sea of gates, it’s more specialized. It requires more software tools and differences in how Xilinx approaches business. It’s a more complex sell. They are pivoting away from the traditional FPGA business.”
Xilinx faces competitors such as Intel and Nvidia on its newest innovations, while it holds a commanding lead over Altera in the FPGA space. Onder said Xilinx recently captured about 65% of the FPGA market. Intel bought Altera in 2015 for nearly $17 billion, but customers routinely refer to its products as coming from Altera, not Intel.
With 5,000 engineers based on every continent and 60,000 customers, Xilinx operates in many verticals and puts a heavy emphasis on products for communications, including 5G, data centre and auto. The auto category puts Xilinx technology to the test for customers doing AI inference work in assisted driving and completely autonomous vehicles.
For example, Xilinx FPGAs and Versal ACAP are used in Level 4 AV systems from Pony.ai to eliminate performance bottlenecks. In another example, the Xilinx Zync UltraScale+ multi-processor System on Chip (MPSoC) is being used to power Baidu’s Apollo platform for an automated valet parking system being used in vehicles rolling out in China, according to the companies.
The Zynq business grew about 70% in the past year, approaching $1 billion in revenue. Xilinx has provided more than 170 million chips to automakers who have 100 ADAS and autonomous vehicle models in production.
While FPGAs are not new, that business segment continues to be robust. “There’s lots of programmable business in hot markets like networking where you may have to change the configuration of networks often to AI where custom algorithms can be programmed into accelerators,” said Jack Gold, an analyst at J. Gold Associates.
Going back to 2016, rumours surfaced that Qualcomm or Broadcom would buy Xilinx for $15 billion. Analysts said it is always possible Xilinx could be purchased, but the ACAP Versal introduction in 2019 seems to have helped dampen those rumors for now.
Xilinx achieved $3.06 billion in revenues in fiscal 2019 and recently forecast a 6% increase in fiscal 2020 revenues despite an expectation of a low point in its third quarter. The stock started 2020 at nearly $100 a share, up from $80 a share a year earlier.
Like many U.S. companies that trade with Chinese companies, Xilinx has been hindered by the U.S. blacklisting of Huawei. “They are an important customer, and we would like to see the U.S. and Chinese government work out an agreement that respects IP rights and enables global trade,” Onder said.
“There are a lot of concerns about the trade war in this industry. China graduates the most engineers in the world every year. They are a vibrant and dynamic economy and offer a lot of innovation. China continues to be an important market for us,” he said.
Tracking future tech
Like other large semiconductor companies, Xilinx has its eye on future technologies, including quantum computing and gallium nitride (GaN) transistors.
“GaN is a direct, wide-band gap material which makes it a viable candidate for light-emitting devices such as LED or laser with shorter wavelength. It might also be a candidate for high-voltage application,” said Xin Wu, vice president of silicon technology for Xilinx.
However, he said there are technology challenges for large-scale integration. Also, Xilinx is fabless and depends on wafer foundries for its semiconductor technologies. “So far, there is not large scale GaN integration available in foundries as we are aware,” Xin Wu added.
Quantum computing is even more speculative for many companies, even though industry giants Intel, IBM, Google, Microsoft and others are researching it. “Quantum computing is still in its infancy. The industry still needs to figure out a comprehensive set of tools for encoding math into qubits and algorithms that exploit the operations which quantum computers provide,” said Ralph Wittig, fellow in the office of the Xilinx CTO.
Considering it took decades to evolve Boolean logic into tools that train and run neural networks, it will take time to retool circuits, tools and algorithms to use qubits, he said. “For now, Xilinx will focus on integrated circuits which follow Boolean math and logic.”
“That said, FPGAs are being used today already to convert Boolean logic into qubit encoding and vice versa,” Wittig said. “Many quantum computers thus use a lot of FPGAs at their interfaces.”
By: Matt Hamblen
You must be logged in to reply to this topic.