IEEE FPGA Engineering Milestone

Tom Coughlin

I recently participated in an event that unveiled a new IEEE milestone for the field programmable gate array, FPGA, at the AMD, former Xilinx facility, in San Jose. [I was president of the IEEE in 2024.] FPGAs were an important technology for speeding important consumer and industrial electronic devices to market. In addition to the event at the AMD facility there was a talk the week before at a Consultants Network of Silicon Valley, CNSV meeting, by Steve Trimberger, formerly at Xilinx about the history of the FPGA.

An FPGA is a semiconductor package that allows using software to create circuitry in a chip. The first Xilinx FPGA, introduced in 1985, was the XC2064, shown below. This chip had 64 flip flops, 58 and operated at 18MHx using 2-micron feature lithography.

FPGAs allows electronic designers to create semiconductor devices that perform various functions much faster and at a lower cost, at least for lower volume products, than application-specific integrated circuits, ASICs. Early ASICS had all of their operations hard wired into the device without any programmability.

As stated in the IEEE milestone site for the FPGA milestone, “FPGAs gave access to leading-edge semiconductor technology for vastly more designers than ASIC (Application-Specific Integrated Circuit) technology could provide. Even with sophisticated Electronic Design Automation (EDA) software, large semiconductor devices are difficult to design as they require a broad range of skills in system design, logic design, circuit design, transistor-level modeling, test program generation, package design, and signal integrity.

As ASIC vendors could not provide sufficient assistance for a large number of customers, small customers were often turned away. By providing a pre-manufactured device, FPGAs eliminated whole classes of problems for their users, and thus facilitated the deployment of custom silicon in a broad range of applications.”

In Steve Trimberger’s talk he pointed out that programmable logic devices, PLDs, had been around before FPGAs but they scale linearly with the complexity of the logic cells. FPGAs logic scaled with process technology and thus Moore’s Law allowed more and more programmable cells in a given size chip. Ross Freeman, inventor of the FPGA and a founder of Xilinx said that, “Moore’s Law is on their side.”

The figure below compares the total costs versus unit volume of ASICS versus FPGAs with additional generations of shrinking features due to Moore’s law. ASICs have a big non-recurring expense at the start to provide all the masks needed to make the chip. On the other hand, FPGAs start out with no up-front costs but the costs per unit increase faster than for ASICs. But with successive Moore’s law generations the number of units where the ASIC versus FPGA costs cross over increases. This is why Moore’s law favored FPGAs.

FPGAs provided a boost to using electronic design automation, EDA, for creating rapid product designs in order to feed device fabrication that spurred new applications and drove FPGA production volume. A particular application that drove FPGAs from 1992 through 2000 was the packet router for Ethernet applications. This drove the move from prior generation FPGAs to platform FPGAs. These devices offered programmability but also included some built-in circuitry for performance standardized functions.

The Virtex-4 in 2005 included a number of built in functions as shown below.

The slowing down of the physical scaling approaches that enabled Moore’s law began to slow down and customers looked for low-cost complete systems built with standard IP blocks. With its Zynq product in 2011 the company introduced field programmable system on chip, SoC, devices with programmable logic that worked with standard IP blocks dominated. In these devices the bulk of the chip incorporated standard IP and libraries and the programmable portion was a much smaller part of the chip as shown below.

Xilinx went through three major changes in their programmable logic products from the first FPGA in 1985 to today. First traditional FPGAs rode Moore’s law to create ever lower cost products with little up-front costs. As Moore’s law slowed, Xilinx began to add products with some built in functions tied together with programmable logic and finally the company came out with programmable SoCs that included a lot of standardized libraries and IP as well as higher level programmable functions to tie all of this IP together for particular applications.

Xilinx was acquired by AMD in 2022, although the acquisition was announced in 2020. Xilinx became part of AMD’s adaptive computing effort to expand into sectors including automotive, healthcare, aerospace and AI.

The IEEE engineering milestone for the FPGA recognized the importance of this device to enable more rapid economical chip design. Over time this programmability was incorporated into SoC designs enabling many modern electronic applications.