[ODE] Havok and NVIDIA present Havok FX at GDC 2006

[Chris Ledwith]

erwin at erwincoumans.com wrote:
>> The hardware is very different. There is not enough conditional flow 
>> control in the ATI/NVIDIA hardware to do "all" of physics (collision 
>> detection; friction models; etc) at a sufficient detail level, except 
>>     
>
> Shader Model 3 gives enough control flow for interesting convex rigidbody 
> physics, not just particle systems: 
>
> An ex-collegue at Havok told me that GPU physics does run the entire physics 
> pipeline (apart from the the broadphase) on the GPU. This includes collision 
> detection, friction and constraint solver. The details about the collision 
> detection are not to be disclosed yet, but it allows for stable stacking and 
> thousands of convex objects on a static mesh. Also Havok GPU physics allows 
> for gameplay physics already, the CPU side can request feedback from the GPU 
> on a per-body basis. So it doesn't require to send ALL the body information 
> (which keeps things fast). 
>
> The Playstation 3 supports both Havok and Ageia accelerated on the Cell/SPU 
> (and potentially GPU), no Ageia add-on. 
>
>
>   
Thanks very much for this info. I'm still dying to know more details and 
comparisons (as I'm sure you are too). There's just not a lot of public 
information yet to form a strategy for the next couple of years.
My impressions thus far are that both technologies (obviously) have 
their advantages and disadvantages. But that the PPU is far more capable 
of churning out physics-relevant (especially highly interdependent, 
difficult to parallelize) calculations in one frame, which makes it of 
greater value (from a technological standpoint), even if the throughput 
back to the CPU turns out to be the same for both techs. Note that a PCI 
express version of the PhysX boards is not yet available but should be 
forthcoming.

I've seen videos of Havok's GDC demos of hardware accelerated physics, 
but of course they were happening in extremely simple 
programmer-art-style environments. I'm wondering about the distribution 
of physics vs. graphics calculations in that setup, and whether a 
single-GPU, single-PPU system would blow it away, even just for stacks 
of bricks and piles of rubble (not fluid physics, for which I imagine 
there's no contest). Of course this is all dependent on the contact 
generation and collision resolution schemes employed too. PhysX seems 
like it could handle greater *fidelity* at the same scale (egads, I'm 
being swallowed by their marketing-speak).

-C

> Jon Watte (ODE) writes: 
>
>   
>> steve1011 wrote: 
>>
>>     
>>> I believe nVidia and ATI are responding to Ageia's h/w by showing how 
>>> physics can be accelerated using GPU's.
>>>       
>> The hardware is very different. There is not enough conditional flow 
>> control in the ATI/NVIDIA hardware to do "all" of physics (collision 
>> detection; friction models; etc) at a sufficient detail level, except 
>> for things like scary-big particle systems. 
>>
>> Graphics really is all about pushing things through a pipe of a 
>> pre-determined length (although the length is re-configurable with 
>> shaders); physics is more interdependent. That's what Ageia has going 
>> for them. That, and the fact that you're not likely to throw multiple 
>> $1,000 CPUs at the problem (if you were using x86 physics), when you can 
>> throw a $200 add-in card at it instead :-) 
>>
>> Cheers, 
>>
>> 			/ h+
>> _______________________________________________
>> ODE mailing list
>> ODE at q12.org
>> http://q12.org/mailman/listinfo/ode
>>     
>  
>
> _______________________________________________
> ODE mailing list
> ODE at q12.org
> http://q12.org/mailman/listinfo/ode
>
>