F1technical.net

Didn't mean it come off so harsh... but I agree with you 100%. As we always tell our customers, crap in equals crap out. Without mesh resolution where resolution is needed, the results will suffer big time.

Unfortunately, for accurate simulations of F1 cars, a full car model is essential as each and every bit has a huge impact on each and every other bit. This requires an epic computer to handle the tens of millions of nodes matrix in a reasonable time frame.

[quote="slimjim8201"]Didn't mean it come off so harsh... but I agree with you 100%. As we always tell our customers, crap in equals crap out. Without mesh resolution where resolution is needed, the results will suffer big time.

Unfortunately, for accurate simulations of F1 cars, a full car model is essential as each and every bit has a huge impact on each and every other bit. This requires an epic computer to handle the tens of millions of nodes matrix in a reasonable time frame.[/quote]

That is EXACTLY what I started this thread to answer. I wanted to know if it was possible to offload that "epic" computing necessity to several 128bit GPU's.

But, whatever, right?

Chris

Just to drag this up...saw these;
http://www.theregister.co.uk/2008/06/17 ... ream_9250/
http://www.tomshardware.co.uk/Clearspee ... 28516.html
http://www.theregister.co.uk/2008/06/16 ... _tesla_10/

...and thought of this thread.

General Purpose GPU's with incredible FPU performance are what you need in a standard workstation.

I'm looking into this at the moment. What kind of performance would one of the Nvida Tesla workstations give compared to well setup Intel or AMD desktop workstation for CFD processing? I'll admit I'm not well read in this subject...

The FireStream and Tesla's are all dedicated maths chips they are made to process a single thread of floating point cals as fast as possible...1TeraFlop is faster than SuperComputers used to make....Infact a single rack of blades with clearspeed cards in them will instantly drop you into the world supercomuter top 250 (but each blade is about £30k).

Either way if you've got a CFD app that can use a Tesla or Firestream then your system will perform 100's of times faster than just relying on a generic CPU from Intel or AMD.

100s might be somewhat overstated, 10-20x is a more realistic number. The problem however, is that your workload must be suited to the chips. For one, it has to be absurdely parallel, we're talking several dozen threads at least to reach full potential. Secondly, you'll prefer a single-precision float situation, as double-precision (which is needed for a lot of scientific work) comes at a 4-8x performance penalty). Thirdly, graphics chips are scarely complex, more so than CPUs, which makes them hard to program efficiently.
What is currently preventing a lot of ISVs to ship for these GPU-based is the cost to convert their apps for this, and the expected cost of having to retune their work for new generations (they can change pretty substantially). Added to that, nVidia uses CUDA as a development platform, ATI/AMD chips have the BrookGPU platform, and Apple is working on OpenCL. As you guessed it, they aren't compatible.
All in all, it's really cool, with some great potential for suitable loads, but it may not make much commercial sense as it stands.

axle wrote:The FireStream and Tesla's are all dedicated maths chips they are made to process a single thread of floating point cals as fast as possible...1TeraFlop is faster than SuperComputers used to make....Infact a single rack of blades with clearspeed cards in them will instantly drop you into the world supercomuter top 250 (but each blade is about £30k).

Either way if you've got a CFD app that can use a Tesla or Firestream then your system will perform 100's of times faster than just relying on a generic CPU from Intel or AMD.

I believe that this is what my original intention was of this thread.

I was willing to build the computer with 8 GPU's and a QuadCore CPU if someone from the community was willing to "massage" a CFD program to harness the GPU's.

Good articles, BTW, Thanks!

Chris