FPGAs Need a New Future

Cost is also such a big issue.

If performant FPGAs were more accessible we’d be able to download models directly into custom silicon, locally, and unlock innovation in inference hardware optimizations. The highest grade FPGAs also have HBM memory and are competitive (on paper) to GPUs. To my understanding this would be a rough hobbyist version of what Cerebras and Groq are doing with their LPUs.

Unlikely this will ever happen but one can always dream.

FPGA toolchains certainly could do with being pulled out of the gutter but I don't think that alone will lead to much of a renaissance for them. Almost definitionally they're for niches: applications which need something weird in hardware but aren't big enough to demand dedicated silicon, because the flexibility of FPGAs comes at a big cost in die area (read:price), power, and speed.

The issue with the software team using an FPGA is that software developers generally aren't very good at doing things in parallel. They generally do a poor job in implementing hardware. I previously taught undergraduates VHDL, the software students generally struggles with the dealing with things running in parallel.

VHDL and Verilog are used because they are excellent languages to describe hardware. The tools don't really hold anyone back. Lack of training or understanding might.

Consistently the issue with FPGA development for many years was that by the time you could get your hands on the latest devices, general purpose CPUs were good enough. The reality is that if you are going to build a custom piece of hardware then you are going to have to write the driver's and code yourself. It's achievable, however, it requires more skill than pure software programming.

Again, thanks to low power an slow cost arm processors a class of problems previously handled by FPGAs have been picked up by cheap but fast processors.

The reality is that for major markets custom hardware tends to win as you can make it smaller, faster and cheaper. The probability is someone will have built and tested it on an FPGA first.

You can pretty much do everything in Vivado from the command line as long as you know Tcl...

Also, modern Verilog (AKA Systemverilog) fixes a bunch of the issues you might have had. There isn't much advantage to VHDL these days unless perhaps you are in Europe or work in certain US defense companies.