Japanese boffins fire up 802 teraflops ceepie-geepie Upstart supercomputer maker Appro International has started up another Xeon E5-based supercomputer, this one a hybrid CPU-GPU Xtreme-X machine installed at the University of Tsukuba in Japan. The deal at Tsukuba was announced last September at about the same time that Intel was widely expected to launch the "Sandy Bridge-EP" …
Suggestion for a future El Reg article:
Do some interviews with the end users of machines like this. I'm sure that 99% of this site's readership are well familiar with regular server operation, but for many, myself included, supercomputers are well in the land of boffinry. We often hear that they're used to do stuff like nuclear weapons research and chemistry. I'd like to hear from the users to describe the class of problems that require such performance, how they went about it, and what some of the results are.
Some interesting stories there along with what a different paradigm these kinds of machines are. For example, 99% load will alert any system admin while on super computers, it means the hardware is fully utilised, not wasted.
www.0x07bell.net/WWWMASTER/CrayWWWStuff/Cfaqp2.html#TOC2
Mostly they feel like a throwback to the days when the dinosaurs ruled the Earth and every computer manufacturer did its own thing.
The comfy world of the GNU tool chain might not be available and in any case you'll need to port your code to use the system provided versions of "foomem" or "barlock" or whatever. And because of that many silly things that one takes for granted might not be at all trivial to port.
The websites for SGI and CRAY have a few documents on that sort of thing. Luckily it is getting better. See, for example, the docs for the COSMOS machine at Cambridge:
http://www.damtp.cam.ac.uk/cosmos/
Yeah I know this has turned into an addicition for me, but if E5's scale anything like the SB desktop chips do under overclocking that that would be quite a large boost in performance :D It would certainly give some of the other supers out there a real run for there money.
The interesting thing is that the stability of alot of these chips today when overclocked and under a controlled temperature and clean power, is almost worth it these days. I am not 100% sure of current lifespan of clusters these days, in deployment requirements and time. But it might just about be a point that its cost effective to do such things. After all there are a few companies selling overclocked servers to wall street. Based around the EVGA Sr2 motherboard and a pair of xeons with bclk cranked into the high heavens.
HPC doesn't generally overclock for two main reasons. first, overclocking is, by definition, running the system outside of spec. unless the specs were stupid, that means less reliable or robust - higher FIT, etc. second, overclocking dramatically increases power dissipation, and operating at scale means optimizing for performance/power, which means a strong preference for lower clocks.
This post has been deleted by its author
another example how intel and nvidia are co-operating to dominate quality chip production
you can see the marriage plainly here .. intel motherboards .. intel bridge chips .. with nvidia GPUs ( onboard? )
intel did not pay nvidia $1.3 billion just to settle IP .. it involved an agreement to stay out of each others' core businesses and use each others' strengths together
nVintel Inside™ .. now that might play Crysis
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
