IMO, everyone is getting more creative since scaling doesn't help performance as much anymore. Moore's Law allowed chip makers to lean on physics to make existing architectures faster. Now the plethora of accelerators (risky new architectures) is proof that physics isn't doing as much of the heavy lifting.

Back in 2013 or so HotChips had a presentation from some dude from DARPA about chip development at and after the end of Moore's law scaling. It probably was the most prophetic tech talk I've ever seen. Your comment could essentially be a quote straight out of it.

Anyway, for me the rub comes in with the fact that historically it's taken a really long time for mainstream software to exploit new instruction sets or architectures.

To be fair, a lot of people had that prediction. But DARPA has definitely been ahead of the game in most research.

More importantly, Dennard scaling has basically already dropped dead since around ~40-28nm, although we did get a "pseudo-continuation" in the form of FinFETs and friends.

Yeah, seems like it's just getting more discrete and less continuous.

That said, this is not an 88x speedup for general purpose computing, just for some specific workload of parallel compute.

If I read this correctly, this is effectively a better way to do thread level speculation (aka Speculative Multithreading), so in theory this could help in general purpose computing as well, just not 88x ;)

Your intuition on this improving Thread Level Speculation (TLS) is correct. I do want to clarify that this is not just a better implementation of TLS though -- our execution model is more expressive and avoids some of the pitfalls of TLS (see the Swarm paper http://bit.ly/swarmarch to learn more). Yes, you could use our implementation with a TLS-like execution model and improve on the TLS performance. But you are likely to reap more rewards by using the Fractal execution model + implementation.

The 88x is the maximum speedup we get, yes. We believe our ideas are applicable to a wide range of domains ranging from databases, to discrete-event simulation, to machine learning, and graph analytics.

Very cool work nonetheless, it's easy for the software folks to feel completely hopeless these days with transistors suddenly getting more expensive per capita instead of cheaper.