Xilinx

» see xilinx.com as it lived, on the wayback machine

~ the last words ~

Of every grave on this site, this is the one closest to my desk. I do edge AI and computer vision for drones in Ankara, and in that corner of the work the question of what the vision pipeline actually runs on comes down to a short list of vendors, and for twenty years one of the names on that list was Xilinx. There is no nostalgia in it. The word sat in datasheets and toolchain installers and part-number fields, and then one year the datasheets started arriving with a different logo on page one and the same part number underneath.

Xilinx was founded in February 1984 in San Jose by Ross Freeman and two colleagues out of Zilog. Freeman's contribution was the idea itself. Until then, if you wanted custom logic you paid for a mask set, waited months, and lived with whatever you had specified. Freeman wanted to ship the customer an array of gates with no function committed to any of them, plus enough on-chip memory to hold a configuration loaded fresh at power-up. Inside the industry the concept was considered wasteful, because most of the die area went to the wiring that made it reprogrammable rather than to the logic doing the work.

The XC2064 shipped in 1985 with 64 configurable logic blocks. It was the first commercially viable field-programmable gate array and it created a product category that had not existed the day before. Xilinx never owned a fab, which in 1984 was an unusual way to run a chip company and is now the normal one. It went public in 1990 and spent most of the next three decades holding somewhere around half of a market it had invented, with Altera taking most of the rest until Intel bought that one in 2015.

By fiscal 2021 the company was doing $3.15 billion in revenue with 4,890 employees, and the product line had drifted a long way from programmable glue logic. Zynq put ARM cores on the same die as the fabric. Versal added AI engines. The parts stopped being components you soldered next to a processor and started being the processor.

~ the record ~

~ the number nobody agrees on ~

AMD announced the acquisition on 27 October 2020 at a headline price of $35 billion, all stock. The deal closed on 14 February 2022. Because AMD shares moved a long way in the sixteen months between those two dates, the price you see quoted depends entirely on which day the writer picked, and figures anywhere from $35 billion up to $60 billion appear in reputable places describing the same transaction. By most of those measures it was the largest acquisition in the history of the semiconductor business.

Victor Peng, who had been Xilinx CEO, stayed on to run what AMD renamed the Adaptive and Embedded Computing Group. He retired in 2024.

~ how a brand gets switched off ~

AMD said from the start that it would not keep the name, which was a different choice from the one Intel made with Mobileye. The phase-out ran through June 2023. Product change notices went out telling customers that the mark on the package was changing from the Xilinx logo to the AMD logo with no change to the die underneath, which is the most bureaucratic possible way for a forty-year-old company to stop existing.

What survived is the vocabulary. Spartan, Artix, Kintex, Virtex, Zynq and Versal are all still on sale under AMD, and engineers still say them in exactly the order and with exactly the meaning Xilinx assigned. The naming scheme is now older than the company that wrote it.

~ what the domain does now ~

xilinx.com resolves. It does not serve anything. Every request lands on the AMD homepage, which makes this the tidiest corporate death I have checked for this site: no parked lander and no squatter selling the name back. Somebody set a redirect, and thirty-nine years went quietly into a marketing group.

AMD has meanwhile been discontinuing the older Xilinx parts on schedule, including the XC9500XL CPLDs and the Spartan-3 family, on the ordinary grounds that the run rate had fallen off and the supply chain was done. That is the second death and the one that actually costs somebody money, because a CPLD designed into a product in 2006 is still in that product now, and somewhere a purchasing department is reading a last-time-buy notice for a chip whose manufacturer no longer has a name.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »