Transformers scale poorly vs. context window size and parameter count.
Which means really impressive when those N’s are small!
I’m but a pundit in this area so don’t know much. But one wonders if there’s a future in burning larger models to FPGAs - whether big enough FPGAs exist (or can be built), and whether locating specialized compute right with the memory it needs can speed things up.
Likely would need a lot of algorithm parallelism work that’d translate back to CPUs/GPUs.
The context window is 16 characters. Talking about tokens per second is meaningless.
its not meaningless. there could be usecases like spell correction.
See also:
https://rits.shanghai.nyu.edu/ai/karpathys-microgpt-on-fpga-...
TL;DR: The CPU implementation was 71x faster than the FPGA.
Note: model has only 4192 parameters.
That post is uninteresting both because they miss the point, and it's not clear a human was even involved to perceive a point to miss. Sure, with an unlimited transistor budget, power budget, and a design clocked at 4GHz fabbed on 5nm one of the best CPU design teams in the world can make a thing that is straight line faster than a one-person project running at 80MHz on a 20 year old 65nm FPGA. Any other answer would be extremely surprising.
Now, there are a bunch of interesting things about this project. Seeing the example of a tiny transformer running on FPGA is informative, and that it was apparently a pretty quick project for one person + robot assistance. Probably some transferable lessons for anyone else doing robo-FPGA development.
yeah, then theres prompt loading too.
but anyone who can fit QWEN-3.6 35B with a sustained ~30 token/s and ~100k context with cache could print money as a hardware vendor.
That just sounds like a 3090.