Re: Nvidia limiting gtx680 compute speed? back

(L) [2012/03/27] [tby Dade] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

>> spectral wrote:About OpenCL performance,
Have you introduce a bug in the partners.nvidia.com web site ? I'm sure that if you introduce an issue and an example (SLG...) they will give you some feedback and improve it !
It isn't a bug or a SLG specific problem, it is one of their latest driver release to have cut the performance of most OpenCL applications. Their OpenCL forum has multiple threads reporting this problem.

(L) [2012/03/27] [tby Dade] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

Blender/Cycles (i.e. CUDA path tracer) users report that the 580GTX is faster than the 680GTX too.
It is just worser at running that kind of tasks; probably by design. Merax's post (with the link provided) has the best explanation I have seen up to now.

(L) [2012/03/29] [tby dsavi] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

>> Merax wrote:I don't think it's a conspiracy, just changing design goals that resulted in less bandwidth per compute unit.
When you see benchmark results like those, it makes you wonder...
I guess I'll be switching to AMD on my new build.

(L) [2012/03/29] [tby toxie] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

Rest assured that there will be something worth the wait on [LINK http://code.google.com/p/understanding-the-efficiency-of-ray-traversal-on-gpus/] (as usual) soon.. As i said earlier, 680 was mainly designed for more perf/watt, especially in games, but it doesn't mean that ray tracing is slow in general..

(L) [2012/03/29] [tby toxie] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

Okay, seems i was wrong with the addy, for now it's only available from here: [LINK http://www.tml.tkk.fi/~timo/HPG2009/index.html] [SMILEY :)]
(and actually the results are quite amazing, too, at least for this kind of pure, raw ray tracing performance!)

(L) [2012/03/30] [tby Vilem Otte] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

Wonder how they actually get that high numbers (noticed they've got code, yummy ... gotta walk through it during the weekend).  On GPU I still get "just" some 10 MRays/s (for primary) in naive (meaning like no packets or stuff) implementation of KDtree ray tracer on Radeon HD 6770 (not the best card, but I think I should get some 5 - 10 times more, still I know that my code is heavily unoptimized) ... ofc I'm just getting into OpenCL (but I smell some sorcery in my code).

(L) [2012/03/30] [tby toxie] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

Its the small details that they have (you really should read the paperand comments on the page), plus i think they have the benefit of the BVH on Kepler.. But i'll have to run some tests myself to judge if othe stuff (kD/BIH) can profit, too.. Plus CUDA, which usually is the better choice on NVIDIA hardware..

(L) [2012/03/30] [tby spectral] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

Interesting,
Does someone test the performance of the Kepler_kernel... on fermi.
And compared the performance of kepler_speculative_while_while.cu vs fermi_speculative_while_while.cu, to check if fermi also benefit of some changes (It seems that they also do some other changes, but not sure they are related for Kepler) ?

(L) [2012/03/30] [tby toxie] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

the other changes are minimal, its mainly changing the amount of tris processed per pass from one to two, and replacing some fminf's with integer variants..

(L) [2012/03/30] [tby madd] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

What's the best performance that can be achieved for the fastest Intel CPUs for AO rays in conference room?

(L) [2012/04/02] [mpeterson] [Re: Nvidia limiting gtx680 compute speed?] Wayback!

>> madd wrote:What's the best performance that can be achieved for the fastest Intel CPUs for AO rays in conference room?
i can just tell you results for fairy. on dual sandy (3.06ghz) around 290mrays/s with ray length of 0.25 * scene-diag.
mp