Status of Physx driver and future support

If it were that simple then x87 would have disappeared long ago.

The SIMD arrays in modern GPUs are comparable to a modern CPU core (in this respect, at least).

 

I never said it was simple. Compilers typically don't support SSE (languages like C++ don't have a native SIMD type), and even if they do vectorizing your code is not trivial. It usually requires writing (inline) assembly to take advantage of SSE.

All I'm saying is that physics is perfectly suited for SSE and a bit of optimization effort would go a long way.

 

Your previous comments appeared to trivialize the matter, that was all I took issue with.

No argument here. I've said several times I believe multi-core processors are very well suited to the task of physics processing.

 

I wasn't really claiming any power advantages, rather a get even. Since the advent of multi-core CPUs, their performance/power/area can increase at the same rate as GPUs. So the gap between them shouldn't widen any more. Wider SIMD units and FMA can even close it a little, while GPUs can't exceed the curve (power being a primary factor). While obviously power is just as important to the CPU, it doesn't stand as much in the way to reach higher throughput. This is all due to CPUs having far lower performance density to start with, so they still suck. They'll just suck less in the future.

 

In the market today, is there an overwhelming need to do physics on the gpu ? Are there physics implmentations in games today that cant be done with a cpu ? As far as i can see with the games I have played so far, it never accured to me that I needed a video card to do physics within a game. I mean, I dont think I missed anything with how games are using physics through the cpu, why would I want it done on my videocard, when All I want from it is to be able to play a certain game with the most settings at high at playable framerates.

 

The point is not to execute the existing calculations faster, but to open up more math power to perform MORE calculations, which allows for more effects and a more immersed gaming experience.

 

You'll generally see the limitation more in games that are CPU bound due to existing physics calculations that might be tailored for a higher power CPU.

Plus, in general devs have "peformance budgets" that they try to stay within so that their game will run reasonably well on most hardware. Physics on a GPU has the possibility of expanding that budget with regards to physics calculations which has the potential of allowing for more realistic rendering of certain animations or other things.

Take Crysis for example, or any game that attempts realistic cloth modeling (as a simple example). Explosions with debris that act independantly and with realistic trajectories and behavior (collisions and resulting new trajectories, etc). Sure it's not absolutely required. But every bit that adds more to the realism in a scene is a plus in my book. Unfortunately too much of it current WILL completely bog down your CPU.

Regards,
SB

 

major update coming






http://www.geeks3d.com/20110511/physx-3-x-roadmap-details-linux-version/

 

Now is the time when quite a few people have a physx capable retired gpu (eg: I have 2, 8800gts + gtx260)
physx could become a lot more popular (i do hope nv allow a nv card doing physx along with another vendors card doing gfx, they did for 1 driver version iirc)
but then again isnt direct compute the vendor neutral way of doing gpu physic's (or maybe opencl)
any plans to allow those to run on a dedicated card ?

 

Still no word on transparent gpu backend implementation (instead of 2 more or less seperate APIs) or cpu multithreading (should be easy in case of the former).

Btw, have we seen ANY gpu-physx titles since Metro2033 and Mafia2?

 

Mafia2 was fizzyX? Really? I didn't even notice.

 

CPU multithreading is already done, has been featured in Mafia2 already. A major rewrite could mean they will actualy start using SSE2 properly.

 

Was it really multithreading in M2 or rather multiple physx solvers running concurrently (one for clothing, one for particles etc)? If the latter, then that was to the best of my knowledge always possible.

 

I dont see anything there that's a direct benefit to the consumer. What about added features and performance?

There's also an updated 2011 poll on middleware popularity with developers. PhysX seems to be doing ok although there's been nothing to shout about on the hardware acceleration front.

http://physxinfo.com/news/5645/physx-sdk-in-top-5-middleware-libraries-used/

 

I can't say I've noticed, and hopefully there won't be any more either. Physx just fragments the market and is nothing but a blatant attempt by Nvidia to corner said market and lock people into their proprietary shit. They were vehemently against this back when 3dfx was still around and now their position has turned 180 degrees - whodathunkit eh!

 

Actualy, Alice: The Madness Returns and Deep Black (an underwater TPS from Russia) will feature GPU PhysX. So it's not quite dead yet. BF3 maybe too?

 

BF3 is more than unlikely, with FB1.5's tight havoc integration, multi player game play dependance, and being a real AAA title (so cost of license / sponsored development means less than having a wider audience).
A license title and a smaller russian dev are much more likely targets (for the bribe)

That's also my impression.

 

ps:
Section 8 prejudice uses physx (gpu?)
although its more of a mission pack/cash-in than a separate game

 

With no mention here (unlike all the (sponsored) 3d vision talk) and especially on geforce.com I'm certain there's no gpu-physx.

 

Its just business and Nvidia got there first, the other company is not really even bothering to do the same level of games promotion as Nvidia does with their programs so its their fault for their loss.