Forty years ago, Xilinx introduced a revolutionary device that allowed engineers to use logic programming right at their desks.
The Field Programmable Gate Array (FPGA) developed by Xilinx enabled engineers to download bitstreams with custom logic to desktop programmers for immediate execution, eliminating the weeks of waiting for chips to return from the foundry. If errors or issues arose, the device could be reprogrammed on the spot.
Kirk Saban, Vice President of Products, Software, and Solutions at AMD, stated in an interview with eeNews Europe: “I have been working in the FPGA field for 27 years, programming FPGAs since 1999.” AMD acquired Xilinx in 2022. “This is probably one of the least understood types of semiconductors; many people know what a CPU is, and under the influence of AI, everyone knows what a GPU is, but very few know about FPGAs.”
The first FPGA chip, the XC2064, was released in June 1985, featuring 600 gates, 64 configurable logic blocks, and a clock frequency of 70 MHz. This was a significant change that propelled the chip into the semiconductor hall of fame.
“When they first started, it was about absorbing logic on the circuit board and providing programmable I/O,” Saban said. “We have come a long way since then, now boasting high-speed SERDES, hardened IP, memory controllers, Ethernet, AI processors, and ARM embedded computing.”
Xilinx was co-founded in 1984 by Ross Freeman, Bernard Vonderschmitt, and James Barnett, who had previously worked at Zilog. Their goal was to use transistor-based programmable arrays instead of gate arrays developed by companies like LSI Logic and VLSI Technology, where transistor arrays were “programmed” with metal layers during the manufacturing process. Xilinx also pioneered a fabless process, manufacturing chips using Seiko Epson’s 2.5-micron CMOS process instead of building its own foundry. Before being sold to AMD, Xilinx closely collaborated with UMC and IBM.
Vonderschmitt served as CEO from the beginning, succeeded in 1996 by Wim (Willem) Roelandts, who joined from HP. Roelandts was followed in 2008 by Moshe Gavrielov, who came from Cadence Design Systems and is now a board member of the company.
“I am honored to have had the opportunity to lead Xilinx for the past decade,” Gavrielov stated when he stepped down as the company’s third CEO in January 2018, with Victor Peng taking over. Peng had previously worked at AMD and was responsible for overseeing the acquisition four years later.
“Xilinx invented the world’s most successful category of programmable logic in 1985 and has maintained its leadership ever since. In recent years, Xilinx has expanded its market share, achieving unprecedented strength, opportunity, and momentum due to the extraordinary talent of our employees,” Gavrielov remarked.
The acquisition in February 2022 made Xilinx part of AMD’s Adaptive and Embedded Computing Group, adding to its embedded x86 processor lineup.
“Some things have changed, but many things remain the same,” Saban said. “We make our own manufacturing investment decisions, and our business units are also responsible for embedded CPUs, Ryzen, and Epyc, as well as a custom ASIC team, so we have evolved from pure FPGAs to embedded x86 and custom solutions, leveraging all of AMD’s R&D.”
One potential advantage of FPGAs is the ability to change the functionality of the device even while it is running. This partial reconfiguration allows for the modification of device modules to replace multiple components, although the process is complex. This has also enabled tools developed by companies like Mipsology (later acquired by Xilinx) to achieve over 100% logic array utilization.
However, even before the recent AI boom, the rise of AI had already provided a rapid growth impetus for Xilinx’s FPGA business.
“Our business units typically operate more at the edge rather than in the cloud,” Saban said. “We have indeed seen a significant shift in edge inference, and our CPU and FPGA technologies are well-suited for real-time processing at the edge. It illustrates the historical advantage of programmability when things change so rapidly.”
Banking and financial institutions became major adopters of this technology in the 2000s through Alveo acceleration cards.
“Fintech is an early adopter market where engineering technologies can leverage real-time capabilities,” Saban said. “They are not using our AI tools but are genuinely using machine learning compilers, writing code at a very low level, so it is more of a silicon architecture game rather than widespread market adoption of edge AI, which requires modern compilers and ease of use.”
At the same time, these devices have attracted the interest of automotive developers for use in entertainment systems and early sensor systems.
“The automotive sector is leading in embedded AI, ADAS, image detection, and low-latency real-time decision-making, with a lot of innovation happening in the automotive field,” he said. “The way cars are manufactured is changing; it is becoming an iPhone on wheels. We started with IVI but have evolved to ADAS and autonomous driving, extending into aerospace as well.”
“There is a huge demand for localized computing with limited power for all those things people are interested in, such as autonomous systems, robotics, drones, and cars, which are very well suited to the products we have. Humanoid robots are also gaining significant attention in many markets, whether in hazardous environments or on production lines. Time to market, field programmability, wireless updates, and other fundamental principles remain very important, and I believe this will not change as we evolve.”
Process Technology
FPGAs are excellent devices for proving new process technologies due to their large number of transistors and redundancy schemes, meaning low yields do not impact device shipments. This helps foundries improve their processes.
However, in recent years, GPUs have been the devices of choice for process validation, while Xilinx has been pushing small chip technology, combining FPGA arrays, high-speed transceivers, and now AI accelerators.
“This has become a cost issue,” Saban said. “We have returned to a certain tier, but we are still exploring 6nm advanced processes, and we are also looking at more advanced nodes; we have work on 2nm. We can leverage all the work AMD has done on GPUs, but only a relatively small subset of customers can afford it in the FPGA market.”
Saban stated, “We have been using various forms of small chips since 2011, and we collaborated with TSMC to develop CoWoS with Virtex-7. FPGAs have evolved to a multi-mode strategy, with 16nm FinFET at the low end, while Versal prioritizes 6nm and transitions to more advanced technologies. This situation has changed over the years. In the past, when you migrated to a new node, you would move all products to that node.”
This also impacts the long-term support of devices. While it may not be the case 40 years later, Xilinx estimates that of the 3 billion devices shipped during that time, as many as two-thirds are still in operation today.
“Our 20nm devices will continue production until 2040, and our 16nm, 7nm, and 6nm devices will be produced until 2045, and we can achieve this on some popular nodes,” Saban pointed out. “For example, for our oldest parts, Spartan 6, we are still in production, and 15 to 20 years later, we will still be using 40nm and 28nm processes, ensuring the shipment of relevant devices like Virtex4 and Virtex5.”
Source | Chip IntelligenceRecommended Reading——
Splitting DCAI, laying off 20%, selling Altera, outsourcing 2nm… Where is Intel heading?
Intel sells 51% stake in Altera at half price
Three companies bidding for Altera
Lattice: FPGA technology accelerates AI inference
A heterogeneous system-level package integrating FPGA and DSP chiplets
FPGA, completely conquered by RISC-V?
Intel’s FPGA division operates independently, reviving the Altera name
Summary of FPGA-based hardware pin assignment design
☞ Business Cooperation: ☏ Please call 010-82306118 / ✐ Or email [email protected]
