AMD celebrates 40 years of Xilinx
Cette publication existe aussi en Français
Forty years ago saw the launch of a revolutionary device that could be programmed with logic at the engineer’s desk.
The Field Programmable Gate Array (FPGA) developed at Xilinx enabled engineers to download a bitstream with custom logic to a desktop programmer to run immediately, without having to wait weeks for a chip to come back from a fab. And if there was a bug or a problem, the device could be re-programmed there and then.
“I’ve been involved in the FPGA space for 27 years, starting in 1999 programming FPGAs,” Kirk Saban, corporate vice president of products, software and solutions at AMD, which acquired Xilinx in 2022, tells eeNews Europe. “It’s probably one of the least known types of semiconductors, people know what a CPU is and with AI now know what a GPU is, but less know about the FPGA.”
The first chip, the XC2064, came out in June 1985, but of course followed many years of research and development as well as design and tapeout earlier in the year. This had 600 gates with 64 configurable logic blocks and ran at 70MHz. But it was a dramatic change that sees the chip in semiconductor hall of fame.
“When they first started it was about sucking up the logic on a board and providing programable I/O,” said Saban. “We have evolved a long way form that today with high speed SERDES, hardened IP, memory controllers, Ethernet, AI processor as well as ARM embedded compute.”
- Farewell Xilinx as AMD deal completes
- Data centre AI drives $30bn XIlinx deal
- Starbleed vulnerability found in Xilinx FPGAs
The company was co-founded by Ross Freeman, Bernard Vonderschmitt, and James Barnett, who had worked together at Zilog, in 1984. The aim was to use transistors with a programmable array, rather than the gate arrays developed by companies such as LSI Logic and VLSI Technology where the arrays of transistors were ‘programmed’ with metal layers in a fab process. Xilinx also pioneered the fabless process, using Seiko Epson to make the chips on a 2.5 micron CMOS process rather than building its own fab. The company later worked closely with UMC as well as with IBM in the days before it sold its fab to AMD.
Vonderschmitt was CEO from the start, replaced by Wim (Willem) Roelandts who came from HP in 1996. Roelandts was succeeded by Moshe Gavrielov in 2008, who joined from Cadence Design Systems where he now sits on the board of directors.
“It has been a great privilege to have had the opportunity to lead Xilinx over the past ten years,” said Gavrielov when he stepped down as the company’s third CEO, replaced by Victor Peng in January 2018 who had previously worked at AMD and oversaw the acquisition four years later.
“Xilinx invented the world’s most successful programmable logic category in 1985 and has maintained its leadership ever since. Over the past years, Xilinx has expanded its market share, achieving an unprecedented position of strength, opportunity, and momentum because of the incredible calibre of our people,” said Gavrielov.
The acquisition in February 2022 saw Xilinx become the Adaptive and Embedded computing group in AMD, adding in the embedded x86 processor range.
“Some things changed but a lot of things stayed the same,” said Saban. “We make our own manufacturing investment decisions, and our business unit is also responsible for embedded CPUS, Ryzen and Epyc and a custom ASIC team so we have evolved from pure play FPGA to embedded x86 and custom and we are leveraging al the R&D from AMD.”
One potential advantage of FPGAs was the ability to change the function of the device even during operation. This partial reconfiguration allows blocks of the device to the changed to replace multiple components, but the process was complex. This also allowed tools developed by companies such as Mipsology in Italy, later acquired by Xilinx, to achieve over 100% utilisation of the logic array.
However it is the rise of AI has been a boon to the business long before the recent boom in AI.
“Our business unit typically plays at the edge more so than the cloud,” said Saban “We do see a big inflection with inference at the edge and our CPU and FPGA technology is well placed for the real time processing at the edge. It speaks well with the historical benefits when things are changing so quickly with the programmability.”
Banks and financial institutions were a key adopter of the technology in the 2000s with the Alveo accelerator card.
“Fintech was an early adopter market with the engineering to take advantage of real time capabilities,” said Saban. “They weren’t using our AI tools rather than really using the machine learning compilers, they were writing code at a very low level so that was more of a silicon architecture play rather than a broad market adoption for edge AI which will need modern compilers and ease of use.”
At the same time the devices were of interest to automotive developers for entertainment systems and early sensor systems.
“Automotive is very much at the forefront of embedded AI, ADAS, image detection, real time decision making with low latency, there’s a huge amount of innovation going on in automotive,” he said. “How a car is being built is changing, it’s becoming an iPhone on wheels. We started with IVI but has evolved into ADAS and autonomous driving and that is carrying through in to aerospace.”
“There’s an awful lot of interest in autonomous systems, robots, drones, cars, all those kinds of things with a huge need for localised compute with limited power envelope, that fits well into the products we have. Humanoid robots are also getting a lot of attention in a lot of markets now, whether in hazardous environments or production lines. The underlying principles of time to market, field programmability, updates over the air, all those things are still very important and I don’t see that changing as we forward.
Process technology
The FPGA was a great device for proving a new process technology as it has a large array of transistors and a redundancy scheme so that low yields did not compromise the shipment of a device. This helped the foundry improve its process.
However in recent years its has been the GPU that is the process proving device, while Xilinx n particular has been pushing chiplet technology to combine the FPGA array, high speed transceivers and now AI accelerators.
“It has become more of a cost issue,” said Saban. “We have come back one notch but we are still working on very advanced process at 6nm and we are working on more advanced nodes and we have 2nm work. We get to leverage all the work the AMD does on GPUs, but it’s a relatively small subset of the FPGA market that can afford that.”
“We have been using chiplets in various forms since 2011,” he said. “We worked with TSMC on CoWoS with Virtex-7. Where FPGAs have evolved to is a multimode strategy, 16nm FinFET at the lower end whereas the Versal premium on 6nm and moving to more advanced technologies. That’s something that has changed over the years. Back in the day when you moved to a new node you moved all the products to that node.”
That also has implications for the long term support of devices. While not for 40 years, the company estimates that as many as two thirds of the 3bn devices that have been shipped in that time are still in action today.
“We have our 20nm parts going out to 2040 and our 16nm and 6 and 7nm parts going out to 2045 and being on some of the popular nodes allows us to do that,” points our Saban. “For example with Spartan 6, our oldest part still in production, we are still shipping on 40nm as well as all of the 28nm devices, Virtex4, and 5 after 15 to 20 years.”
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
