bit-tech Richard Huddy Interview - good Read

[Squilliam]

It's worse than that. Toss an NV PhysX gpu in that competing vendor rig and PhysX will run fine, but AA is still disabled unless you hack the inis to fool the proggy into thinking it's an NV card and then, miraculously, AA works.

The "Nvidia workaround" was to force AA through CCC which gave very slow AA compared to the in-game MSAA. Nice trick to win the benchies.

Considering the fact that the test system was a Core i7, I cannot see how a processor with that much floating point performance could get bogged down so badly by PhysX. I've heard that PhysX is code which is copyrighted, I wonder how that effects the ease of a CPU implementation. Its like the game turns the CPU into a single core P4 2.4Ghz Northwood once its activated. Surely the out of the box implementation should at least use two cores standard!

How can developers even take it as a serious Physics implementation if its only good for technology demos on the latest Nvidia GPUs whilst none of the older G92 and below can implement it without taking a massive hit to performance and none of the ATI cards can do it at all if they don't have a decent CPU implementation?

 

[MfA]

Free license and free developers which will also help for your console versions, where the PhysX probably isn't quite so atrocious (not to mention the physics budget is smaller on consoles anyway, so a pure single threaded implementation might be enough if it at least has some decently optimized SIMD code ... unlike on x86). It's a pure business decision ...

 

[AlNom]

http://pc.watch.impress.co.jp/img/pcw/docs/343/671/08.jpg

http://pc.watch.impress.co.jp/img/pcw/docs/343/671/09.jpg

http://pc.watch.impress.co.jp/docs/column/ubiq/20100121_343671.html

Someone want to explain why G16R16 format is relevant here MSAA did work with Gears of War...

 

[Lonbjerg]

Considering the fact that the test system was a Core i7, I cannot see how a processor with that much floating point performance could get bogged down so badly by PhysX. I've heard that PhysX is code which is copyrighted, I wonder how that effects the ease of a CPU implementation. Its like the game turns the CPU into a single core P4 2.4Ghz Northwood once its activated. Surely the out of the box implementation should at least use two cores standard!

How can developers even take it as a serious Physics implementation if its only good for technology demos on the latest Nvidia GPUs whilst none of the older G92 and below can implement it without taking a massive hit to performance and none of the ATI cards can do it at all if they don't have a decent CPU implementation?

I got this link somewhere on this forum:
http://golubev.com/about_cpu_and_gpu_2_en.htm

Only it gets worse with physics, as physics scales badly (and not at all linear) with more CPU cores.

Compare it to running the graphics on the CPU...it the same thing, people should stop thinking the CPU is better than it really is...because it isn't.

If we alter you statement to:
"Considering the fact that the test system was a Core i7, I cannot see how a processor with that much floating point performance could get bogged down so badly by rendering graphics"

Do you get it now?

 

[aaronspink]

I got this link somewhere on this forum:
http://golubev.com/about_cpu_and_gpu_2_en.htm

Only it gets worse with physics, as physics scales badly (and not at all linear) with more CPU cores.

Compare it to running the graphics on the CPU...it the same thing, people should stop thinking the CPU is better than it really is...because it isn't.

If we alter you statement to:
"Considering the fact that the test system was a Core i7, I cannot see how a processor with that much floating point performance could get bogged down so badly by rendering graphics"

Do you get it now?

Physics scales badly with more CPU cores? Um... Someone might want to do some actual research instead of just making things up on the fly. Physics is EASIER to scale with more CPU cores than it is to make it run fast on GPUs. Here's a newflash, the majority of all physics in the world is run on massively multi-processors CPUs. There's a reason by governments spend upwards of 1 BILLION on by supercomputers that are effectively MPPs of OTS cpus.

 

[Mize]

My point was that even if PhysX scales poorly on i7, NV still clearly haves Batman:AA so that ATI cards would be hobbled. It's no different than the renaming exes of old tricks except this one's harder to detect & defeat.

Even if you do fool the game into thinking your cypress is NV & hacking you NV drivers to allow PhysX to run on your second NV gpu, the proggy will re-detect you card he next time and disable in-game MSAA that was previously working.

I think NV makes great products, but their "marketing tactics" suck.

 

[entity279]

(this post continues on what aaron said)

That's why I am not impressed at all by nV PR on this one.

If they really say that GPU is better for Physics then they should just point us to a few algorithms that are implemented on the Physics API and are often called on current Physics implementations. I'm sure such algoritms are rather well known and it's easy to determine which implementation (CPU/GPU) would be better.

It was said (forgot the links, but surely have been posted on this forum) that some functions have GPU only implementations. Ok, than why did they claimed that is up to game devs to implement multicore CPU Physics if the CPU part of the API is incomplete?


Now really, i just don't like their "we're the good guys, we give Physics to everybody (but nobody=AMD wants it)" attitude. If they want to increase GPU sales with Physics help, fine by me. But then just say something (as PRish as it is) like "we will concentrate on Physics for nVidia GPU, because we believe they are the future of gaming"

 

[MfA]

Havok has multithreading, Bullet has a multithreaded solver (no broadphase yet though), that Infernal Engine is obviously multithreaded, a quick google found BEPU-Physics with a multithreaded broadphase and solver.

 

[Lonbjerg]

Physics scales badly with more CPU cores? Um... Someone might want to do some actual research instead of just making things up on the fly. Physics is EASIER to scale with more CPU cores than it is to make it run fast on GPUs. Here's a newflash, the majority of all physics in the world is run on massively multi-processors CPUs. There's a reason by governments spend upwards of 1 BILLION on by supercomputers that are effectively MPPs of OTS cpus.

I was talking milk-of-the-run desktop PC's, not server clusters, feel free to prove me wrong

 

[aaronspink]

I was talking milk-of-the-run desktop PC's, not server clusters, feel free to prove me wrong

You do realize its easier to extract parallelism with something that is using a big large cache with low latencies for communication rather than message passing over less than 1 GB/s of network bandwidth with multiple millisecond latencies, right?

So, at this point, I think I've done more than enough shooting fish in a barrel over your hypothesis, maybe you should try actual proof for your preposterous claim. As far as my hypothesis, their are decade of research available including multitudes of millions of dollars spent in both software and hardware. I think all the proof I need is readily self evident (hell if it parallelizes on a GPU, then its trivial to parallelize it on a couple cpu cores.)

So you can either admit you were wrong or admit that you are either a paid shill or a troll. Which will it be?

 

[willardjuice]

I was talking milk-of-the-run desktop PC's, not server clusters, feel free to prove me wrong

So PhysX can be parallelized over a GPU, but not a CPU? I don't get it.

 

[AlNom]

So PhysX can be parallelized over a GPU, but not a CPU? I don't get it.

Unbeliever!

 

[entity279]

So PhysX can be parallelized over a GPU, but not a CPU? I don't get it.

aaronspink didn't say it can't be parallelized on GPU so what are you talking about?

hell if it parallelizes on a GPU, then its trivial to parallelize it on a couple cpu cores

And it's not just about Physics. Is about physical simulations done on a computer. They've all worked on the CPU in the past, so why should they be so under-performing now?

 

[willardjuice]

aaronspink didn't say it can't be parallelized on GPU so what are you talking about?

And it's not just about Physics. Is about physical simulations done on a computer. They've all worked on the CPU in the past, so why should they be so under-performing now?

?

I was (I thought clearly) replying to Lonbjerg. I don't even know how you thought I was responding to Aaron.

Unbeliever!

What!?! Everyone knows I think PhysX is bigger than 3D!!!111

 

[trinibwoy]

And it's not just about Physics. Is about physical simulations done on a computer. They've all worked on the CPU in the past, so why should they be so under-performing now?

Underperforming compared to which other CPU based physics engine?

 

[Groo The Wanderer]

(this post continues on what aaron said)

That's why I am not impressed at all by nV PR on this one.

If they really say that GPU is better for Physics then they should just point us to a few algorithms that are implemented on the Physics API and are often called on current Physics implementations. I'm sure such algoritms are rather well known and it's easy to determine which implementation (CPU/GPU) would be better.

It was said (forgot the links, but surely have been posted on this forum) that some functions have GPU only implementations. Ok, than why did they claimed that is up to game devs to implement multicore CPU Physics if the CPU part of the API is incomplete?


Now really, i just don't like their "we're the good guys, we give Physics to everybody (but nobody=AMD wants it)" attitude. If they want to increase GPU sales with Physics help, fine by me. But then just say something (as PRish as it is) like "we will concentrate on Physics for nVidia GPU, because we believe they are the future of gaming"

If NV actually wanted to show how much better GPU physics were than CPU, instead of putting out a hobbled versions that runs x87 code instead of SSE/2/3, they would put out something that is decently optimized for the CPU and scales well. Then they could simply say, assuming that they are not wrong about the relative speeds, and they are, that you need to have (pick a number) 12 cores before a Nehalem is faster than a 9800/250/350/450, and can prove that with scaling. The way they do it now just makes them look like weasels.

And I also agree that the best tool they have to fight Intel right now is physics, but they are blowing their own legs off with a shotgun, limping around and looking for more ammo to remove those pesky knees with instead of working WITH ATI to go after Intel. Then again, NV has proven to just about everyone that they can not be worked with, so no loss in reality.

NV management is playing this totally wrong, but is that news?

-Charlie

 

[entity279]

?
I was (I thought clearly) replying to Lonbjerg. I don't even know how you thought I was responding to Aaron.

Indeed. Let's just say that my axon pattern is different than yours . That and I was quickly reading through the last 2-3 posts .

Underperforming compared to which other CPU based physics engine?

So the other (less used) physics engines for games are just as bad on CPU?

 

[Tamlin]

That makes Batman even weirder. Its the only game I've seen where there is no performance hit going to 2560x1600. But, if Sontin is right and its the threads that are starved, shouldn't that effect the framerates more on a larger resolution?

I decided to try it out with Batman myself and it looks like its still weird. On a 5870 (vs. 4870 in Tomshardware review) with a Q9550 (vs. I7 920 in Tomshardware), I got almost the same in Batman benchmark going from 1280x768 to 1920x1200 with same settings as tomshardware. 1 FPS performance hit in average:

1280x768:

1920x1200:

 

[Sontin]

You get the same fps because your cpu can't calculate the physics information fast enough. Your graphics card is waiting for work. It's the same with the other cpu cores.

 

[Tamlin]

You get the same fps because your cpu can't calculate the physics information fast enough. Your graphics card is waiting for work. It's the same with the other cpu cores.

I was just about to edit my post. I ran another benchmark and this time with PhysX on normal (not off). Still the same average, though maximum went higher: