more than 50GFlops on laptop

[Michael Creel]

I've been playing around with the NVIDIA Cuda SDK, just ran some code that got more than 50GFlops on a laptop. Pretty impressive!

michael@yosemite:~/NVIDIA_CUDA_SDK/bin/linux/release$ ./nbody -benchmark -n=2000
Run "nbody -benchmark -n=<numBodies>" to measure perfomance.

2000 bodies, total time for 100 iterations: 132.397 ms
= 3.021 billion interactions per second
= 60.424 GFLOP/s at 20 flops per interaction

[sudhang]

Very interesting!

Which graphics card does this laptop have?

[Michael Creel]

It's a Macbook Pro, and the card is a 9600M GT. Incidentally, Kubuntu 8.10 works very well on the MBP.

[vk2emq]

Michael,

you've done what I was thinking about in my post of "Pelican next generation - using GPUs ?" - vk2emq Oct 13, 2008

Would it be possible to run CUDA for Linux on a cluster of PCs/Laptops with NVidia GPUs running Pelican - as per thought of the day below?

The FASTRA project from the University of Antwerp gives inspiration to what can be achieved on PC based equipment see

University of Antwerp makes 4000EUR NVIDIA supercomputer

http://www.dvhardware.net/article27538.html

Researchers at the University of Antwerp in Belgium have created a new supercomputer with standard PC gaming hardware. The system uses four NVIDIA GeForce 9800 GX2 graphics cards, it costs less than 4000 euro to build and runs NVIDIA's CUDA technology.  The researchers ran some benchmarks and found that in some cases their 4000 euro desktop superPC outperforms CalcUA, a 256-node supercomputer with dual AMD Opteron 250 2.4GHz chips that cost the University of Antwerp 3.5 Million euro in March 2005.

Project Web Site

http://fastra.ua.ac.be/en/specs.html

Issues re FASTRA
-  cons
use of 64 bit WinXP, heat generation of video card array, lack of redundancy in hardware platform, performance limited by max. of 8 GB RAM on motherboard

- pros

performance, affordability - base hardware - motherboard, graphics, CPU, RAM all inexpensive COTS, open source of CUDA - also available for LINUX

Thought of the day

Could performance be further enhanced by having an HPC of PCs/Laptops with GPUs - providing
redundancy,
more total RAM,
fewer GPUs per motherboard - better heat dissipation and greater choice of motherboard/laptop,
smaller power supplies,
better potential for overclocking each PC,

- and running CUDA for LINUX on top of Pelican ?

[Michael Creel]

I'm sure that you could use CUDA on a Pelican cluster, but how to do so is way beyond me at the moment. I'm now playing with pycuda, and learning how to do simple stuff on a single machine. Taking advantage of multiple GPUs seems like the next step, but I'm nowhere near doing that yet.

I have no plans to try to integrate CUDA stuff with Pelican. I currently view GPUs as a substitute for a cluster, rather than as a complement. Maybe in the future new CPUs will have GPU-like capabilities, and it might get easier to use this sort of technology on a cluster.