Performance scaling : Newton Game Dynamics

[fraggy]

Just wonder if Newton Game Dynamics are optimize for multicore ?
Intel will release 8 core processor during this year (even an 8 core HT) and we expect 32 core in 2 to 3 years from now.
Does NGD scales on 32 core HT processor ? (with 64 threads)

Vincent

[Julio Jerez]

newton scale to an unlimited number of cores.
I does that already.
the biguess problem is not newton is the high cost of thread tasking.

[fraggy]

the biguess problem is not newton is the high cost of thread tasking.

I might help on the "cost of thread tasking" part : what is the actual performance gain on a quadcore for NGD ?
->I can actually compute x2,8 more objects on a quadcore (even x3.1 more if the application priority is set to HIGH)

[Julio Jerez]

actually in a quad core you get a net lost with large scnes.

on dual cores is about 95% almost 100% depending of teh kind of scene.
on the new intel nahelalen when no using hypertreading yidl is mopre tha 90% on each core, with hpetyreading no gain at all, in fat a net lost.
it have nothing to do with settiong thread priority high or low, in fact it shoul be set to the lower possible or what ever the sytem deside, teh problem in teh architerue, intel CPU are no designed fo lwo latency task swithig.

I will way for Larrabe which I think will smoke all the GPUs, until then dual cores is the best bet for Netwon, and for people with Nahelan 3 or 4 thread is the best option.
For people with Quad core I recomned using only two threads, using three or four is a bad because they are not really muticores and memory have to go trough the bus killing the performance.
Netwon really stress the CPU.

you can test it youself,
viewtopic.php?f=9&t=4922

[fraggy]

on dual core is yet is abpy 95% to 100%

You mean 1 thread for newton, and 1 thread for the rest (3D/gameplay/IA...) ?

I got a 4x speedup on a Quadcore HT configuration (corei7 2.6Ghz) I can deal with 10k sphere compare to the 2.5Ksphere with 1 core. Do you think NGD (or any other physics engine) may need that kind of speedup ?

[Julio Jerez]

That sound very good, I am positive that some open source tecnology could used that kind of speed up.
for now I am happy wih the result I am getting. Thank you anyway for the offer.

here is a video of an old scene of newton running in a quad core AMD with more than 30000 bodies, but those are not spheres they are boxes.
http://www.youtube.com/watch?v=L-0iCSve ... annel_page

had you at least played the demos of the SDK, see how they feel, in your IntelCore7?

[fraggy]

here is a video of an old scene of newton running in a quad core AMD with more than 30000 bodies, but those are not spheres they are boxes.
http://www.youtube.com/watch?v=L-0iCSve ... annel_page

I'am still impress when I see that kind of video
I'am not a physics developper, sphere is so simple to collide, it's barely the only object i have in my test demo.
30K cube is impressive !!!

I'am very interessted in performance scaling, have you conduct some tests scaling ? from 1 to 4 threads (on quadcore) with fixed frame rate on realtime (demo should be execute within a given amount of time)

[agi_shi]

Like Julio said, download the latest SDK. Inside you'll also find the compiled demo (which contains many, many different scenes), check it out (it has a GUI and you can select the amount of threads that you want to use, etc.). There's some high-number stacking examples inside the demo.

[fraggy]

thank you, I will do that.

rgds,
Vincent