The FPGA inside a DMA card

Open the listing photo and every DMA card in this market is a fairly ordinary object. An FPGA, a USB 3 bridge, a flash chip, a PCIe edge connector, and a few passives. The FPGA is almost always from one family. The part number printed on it is where most of the market's confusion about "which card is better" comes from. It is worth ten minutes to learn what that number is, because it turns a shopping choice into an engineering one.

Reading the part number

Take XC7A100T. XC is Xilinx's commercial prefix. The 7 is the 7-series generation. The A is Artix, the cheap, low-power member of that generation, as against Kintex or Virtex. The 100 marks relative logic density. It is not a count of anything directly. The T means the part carries GTP serial transceivers, which is how it can talk PCIe at all. So the T is not a size unit. "A 100T" is short for one part in a family, not a tier of quality.

Figures from the AMD/Xilinx 7 Series FPGAs Data Sheet: Overview (DS180). The open pcileech-fpga designs target parts across this range.

What a bitstream is

An FPGA is not a processor running a program. It is a field of logic blocks, lookup tables and routing. A bitstream is the setting that decides what those blocks become and how they connect. Load one bitstream and the silicon is a PCIe endpoint. Load another and the same silicon is a logic analyser front end. The design is written in a hardware description language. Vendor tooling then synthesises, places and routes it, and the output is the bitstream.

These parts hold their setting in SRAM, and that has a consequence worth knowing. The setting is volatile. It is loaded again from an external flash chip at every power-up. That is why a card has a flash chip and its own flashing path. It is why a bad write can leave a board that powers but does not enumerate. And it is why "flashing firmware" on a DMA card is a different job from flashing a microcontroller.

Why the same family shows up everywhere

Artix-7 is a cheap, general-purpose FPGA with PCIe and serial transceivers built in. That pairing is useful anywhere you have to move a lot of data between a bus and something custom. So it turns up in fields with nothing else in common. FreeSRP, an open-source software-defined radio platform, runs its gateware on an Artix-7 XC7A50T. The same family is standard in FPGA development boards and in PCIe prototyping cards. The Xilinx AC701 evaluation kit is an Artix-7 board, and the open pcileech-fpga designs support it.

This matters for one reason. An Artix-7 on a PCIe card is not exotic, and it was not built for this. It is a stock part with a large legal market. That is why the hardware is cheap, why supply exists at all, and why the real differences between boards are almost never the FPGA.

What to actually decide on

1. 1Whether firmware is published for your exact board, not just your FPGA class. A board nobody builds images for is an ornament.
2. 2The USB bridge and which socket is the data port. This costs more builds than logic capacity ever will.
3. 3Fit and clearance in the machine it is going into, measured rather than guessed.
4. 4Thermals, particularly on M.2 parts in a cramped slot with no airflow.
5. 5Logic capacity, last, and only once the three items above are settled.