Neelesh is my childhood friend. We recently got connected after some 20 years – thanks to Web2.0… To my last post “High Perfromance Computing – not the same anymore”, he made a very interesting comment and it cannot stay tucked inside the comment corner of my blog but be up here for everyone to read. Neelesh is now based in NY and works for Citibank as Vice President (I can now guess VP for what?).

Thanks Neelesh!

====================================

Uttam,

Its interesting you mention about CUDA and the advent of GPU based computing.

This is something most of the high-end investment banking houses here in NYC and also in the UK are seriously looking at. We currently use symphony based grid orchestration for lot of our compute power needs for various risk calculations (monte-carlo simulations, prepayment modeling and fixed income risk calculations).

But now several teams are trying to evaluate moving to GPU based computing. My initial reactions to this was that GPU based computing is good if you are doing a “pure-maathematical-grind” type of calculations rather than something that requires an “if-else” type structured logic. Well that does not come to me as a surprise as GPUs were primarily built for that kind of stuff – image rendering or graphics type programming, where matrix algebra is the cornerstone for all computations.

What we were a little hesitant as part of our evaluation and future possible adoption was that someone using CUDA is sort of married to the architecture and its hard to unplug out of it. Most of the code logic in CUDA world is very intrusive and not platform agnostic. Might not be a big issue but this is something software engineers are usually cautious about. (P.S: we were looking at the NVIDIA based products).

But on the flip side there is huge-huge cost savings in hardware needs and data center needs. We were amazed by the amount of reduction we would have in h/w if we ever redid all of our code using the CUDA based API.

While this space is sure to evolve in the very near future in terms of compute power needs, I also have a strong suspicion that the likes of Intel/AMD will try to put something in place in their architectures to prevent or limit folks from leveraging the massive GPU power or at least make sure that they are somehow the part of the equation. At the end of the day the “instruction-set” or the brain power still is with the main CPU and they (Intel/AMD) could tweak things around…

Cheers!!

====================================

Scientific and Enterprise computing that are CPU intensive is not going to be same anymore. GPU architectures are making those large G/T/P FLOP systems faster, greener and smaller, also cheaper.

More and more applications are getting accelerated with CUDA. Among good examples are – CFD solvers running at 1/20 of resource for a given iteration; HMMER running at 62x time faster; Monte Carlo Simulations running up to 50x time faster.

The GPU architectures from CUDA and FireStream have already made it to Top500 list sometime ago which has been a milestone and is fast moving up the charts, I am sure the 35th list of Top500 is going have much more of GPU+CPU. It is only matter of time that desktops would be replaced with supercomputers. The GPU’s help offloading non-sequential parts of application from the CPU and achieves performance of nature described before.

More interesting to watch is if CUDA is really breaking the dominance of Intel and AMD? And if more and more developers will optimize applications with OpenCL or DirectX for CUDA? What will even happen to ClearSpeed?

Cheers!

Uttam