I was … but then it was determined that I needed to be giving 2 of the STAC summit talks in Chicago and NY on the day jobs’s systems.
Then this week, Tianhe-2 info came out, and … well … WOW!
Great job guys! (ob-day job: “could we interest you in some monsterously fast storage to go with that space-time fabric warping super?")
I was speaking recently with a VC we had talked to in early 2002 about accelerators. Vipin and I marveled at how good our predictions were. We never pitched these guys the accelerator side. I wonder for those whom we did pitch, if they are looking at this large, growing, disruptive multi-billion dollar market and shrugging as they fund their next social web 10.0 company. I won’t name names, but it shouldn’t be a badge of honor that all of them missed this market (and considering how dead on we were …).
Tianhe-2, like Titan, is an accelerated supercomputer. It was our belief that the entire market would be shifting to accelerated systems in the 2013-2016 region, with a very small number of non-accelerated systems remaining. We viewed that as a creative destruction event.
One of the primary arguments we had made for accelerators, and APUs in general (before AMD co-opted the term to give it a slightly different marketing bent), was that conventional CPUs were not trying to maximize the computational cycles for particular problems per unit time. Vector and parallel Vector architectures were needed for realizing parallelism, and easy mechanisms of “drop in” style programming were required … libraries, etc. You need this parallelism, and it has to be easy to use or adopt, in order to win converts. It has to be economical, and advantageous … 5-20% performance advantage is great if there’s effectively no work involved. 5-10x performance advantage (wall clock time for whole app), with little cost, and minor/moderate effort is a win as well.
Basically for something like this to win, it has to be obviously a win. You need the performance to be there, and realizable to mere mortals. You need the costs to be reasonable so people are willing to give it a shot.
Forcing people to recode apps is a losing battle. Its been so in the past, and will remain so in the future. Massive momentum requires a massive force to change. No one will rewrite apps from the beginning, before trying simpler drop in style experiments.
Those experiments are now in the past. Most everyone gets the need for accelerators. I suspect that adoption will hasten, and we will see accelerators, and high performance clouds (public, private) start up as well.
Stay tuned on that latter bit. We might have something to say there :)
This said, ISC13 is going to be very interesting. Vipin, Russell, Doug, and Ken are going for us, so if you are there, please stop by our booth and say hi to them. I will be in Chicago and NYC giving some lightning talks. Though I’d love to hear more about Tianhe-2 (so if any of the folks who built it or programmed it are reading this, please send notes, pointers, articles, etc.). I’ll follow Rich B’s coverage at InsideHPC.com and ask the team to take pictures and post to our blog.
Maybe next year …