Folding Forum

Computing can be fast, but majority of time is being wasted by sending bits between various chips within the system and outside of it. More improvements are needed for interconnections.
Maybe it will prove fruitful: http://www.intel.com/content/www/us/en/ ... video.html

Ivoshiee wrote:Computing can be fast, but majority of time is being wasted by sending bits between various chips within the system and outside of it. More improvements are needed for interconnections.
Maybe it will prove fruitful: http://www.intel.com/content/www/us/en/ ... video.html

Though, I think that the whole idea of the F@H project (distributed computing in particular) sort of tries to make that less relevant by making each part of the computation as independent as possible.

SodaAnt wrote:

Ivoshiee wrote:Computing can be fast, but majority of time is being wasted by sending bits between various chips within the system and outside of it. More improvements are needed for interconnections.
Maybe it will prove fruitful: http://www.intel.com/content/www/us/en/ ... video.html

Though, I think that the whole idea of the F@H project (distributed computing in particular) sort of tries to make that less relevant by making each part of the computation as independent as possible.

Exactly. The downside is that there are certain limitations on the simulations that can be done (e.g. extremely large proteins or proteins in complex media, etc), but the upside is that it's not limited by proximity, so each processing unit can be anywhere in the world. Of course Folding@home has another advantage not related directly to the architecture of distributed computing that Stanford manages to take advantage of, that being that it can take volunteers more easily. People seem to be more willing to donate directly by volunteering idle CPU time in a way that they can watch the progress on their own computer rather than donating indirectly by paying money which may or may not go to parts for a potential supercomputer.

I saw an article about optical inter-chip communication probably back in 2009 or thereabouts. So this is not exactly new news for me, but it's still promising. I'll get more excited about it when I see it in consumer- or server-grade hardware.

If interested - video recordings of sessions from the NVidia GTC conference 2013:

James Phillips: Petascale Dynamics on Titan and Blue Waters
http://www.gputechconf.com/gtcnew/on-de ... =&select=+

Vija Pande: Folding at Home: Petascale Scientific computing
http://www.gputechconf.com/gtcnew/on-de ... =&select=+

Scott LeGrand (Amazon): From Folding@Home to AMBER: Five Years of Molecular Dynamics with CUDA
http://www.gputechconf.com/gtcnew/on-de ... =&select=+

A nice overview on the broader challenges ahead - The Road to Exascale. A bit dated, but still mostly valid
http://science.energy.gov/~/media/ascr/ ... report.pdf
A bit more detailed:
http://users.ece.gatech.edu/mrichard/Ex ... 100208.pdf

From the recent ISC conference
Bill Dally, NVidia: The road to Exaflops
http://www.hpcwire.com/hpcwire/2013-04- ... uting.html


On the challenge to not only write portable scientific code, but to preserve performance in portable scientific code
Improving Performance Portability in OpenCL programs, Univ Chicago, ISC 2013
http://hgpu.org/?p=9079 (download the pdf file)

For those who are curious about the future of Folding@home, the last half of my NVIDIA talk addresses some of those aspects.

http://www.gputechconf.com/gtcnew/on-de ... =&select=+

Has there been more improvement in these fields in the past 5 years?

JonazzDJ wrote:Has there been more improvement in these fields in the past 5 years?

Nvidia has been working on faster interconnects than PCIE, largely for super computers.

https://en.wikipedia.org/wiki/NVLink

https://i2.wp.com/timdettmers.com/wp-co ... rmance.jpg
From http://timdettmers.com/2017/04/09/which ... -learning/