AI driven HW/SW CoDesign

Kevin Cameron

For anyone interested in building flows to get from C/C++ to hardware, the key piece is having very fast interactive simulators that can model hardware well enough to judge whether experiments are better or worse than each other. A key piece of that is knowing what the power consumption is, so digital simulators won't do the job. Fortunately there is a high performance analog simulator for free -

You might think an analog simulator is too slow, but what makes them slow is interaction between the models. One can applying AI techniques to transistor level models to generate block-level behavioral level models which perform at O(n) rather than O(n.log(n)) speed since they don't need the matrix solving. Methods are in here -

Adaptive Inverse Control, A Signal Processing Approach: Widrow, Bernard, Walach, Eugene

Since that technique creates models as (event driven) neural-networks, you can use AI hardware as an accelerator for the simulation, so simulation time will come down by the number of cores applied.

The other piece is making the the simulator interactive so that small tweaks to designs don't require a rebuild/restart, that's fairly feasible with the discrete models and the fact that the simulator is open-source.

Ping me if you want to build it :-)