That's nonsense. PCI/PCIe FPGA accelerator boards have been around for ages, mostly used by the finance traders and for machine learning/math-intensive computations before cheap GPUs with much simpler programming models (no need to write e.g. matrix multiplication in VHDL/Verilog anymore) like the various programmable shaders, CUDA and OpenCL have pretty much relegated them into obscurity.

FPGA is great if you need to talk to some hardware very fast/on many pins. E.g. something like a network router where you are shuffling packets between many high speed interfaces. Or doing a lot of measurements/interfacing a bunch of high speed sensors.

But not for general purpose software - GPUs are both faster, easier to develop for (and with good tooling) and much cheaper for doing that today.

Looking forward to that, without 'you need to buy Skylake CPU that is just as fast as the one you had before to get H265 decoding'. Good riddance.

The other thing is that they're a nightmare to programme and they're expensive!

Yes, but speedup can be huge. For example, for NFA processing, using (a couple of) 2W FPGA against 200W GPU "GPUs underperform FPGA by a factor ~80-1000x across datasets" http://people.cs.vt.edu/~xdyu/html/ics17.pdf

The state transition table is encoded in logic for FPGA and as global memory table for GPU. Why didn't they try to encode state transitions as code on GPU?

Also, did they try to use enhanced locality introduced by processing several streams on GPU? E.g., if you keep states sorted as for tuple (state id, stream id) for all your streams, you may get more memory-controller friendly access pattern. I haven't seen mentions of that technique (which MUST be considered after Big Hero 6 [1] - they used that technique to never miss caches in whole movie rendering process). Big Hero 6 is 2014, the paper is 2017.

[1] https://en.wikipedia.org/wiki/Big_Hero_6_(film)

I really do not like papers like one linked by you. One system gets all of the treatment while other ones get... whatever is left.

I guess have they tried to use these techniques for GPUs, they would get performance gap that is much less than reported.

Today they are. Programming can be made easier with large frameworks. Cost can be reduced with higher volume.