DescentBB

clicky

It's not a long read and cuts to the chase. Very interesting. It looks like Mobo's will need to change a bit to accomodate this as well.

This reaffirms my belief that buying a dedicated physics card is just plain dumb: in 12 months all GFx cards will process far more physics than an Agiea card.

I wonder how many of these physics cards will sell? I'm thinking less than 100,000 - and most of those bundled in pre-built gaming systems.

It sounds more like it the way the industry itself is headed so I would imagine it you won't have a choice. It will be in all the GFX cards. *shrug*

PC's are beginning to look like a baddly gabled house. they need to tear it down and rethink them. Easier said than done of course as thry to set up standards in a process like that would take a decade.

Right now I would say the odds are dedicated physics processing hardware will work it's way directly into the GPU core within a couple years. I'm expecting a race not unlike the shader race.

A query from someone not familiar with the specifics of computer architecture: why would physics calculations be handled by the GPU instead of directly by the CPU?

Top Gun wrote:A query from someone not familiar with the specifics of computer architecture: why would physics calculations be handled by the GPU instead of directly by the CPU?

because it can. the physics model gets updated frame by frame anyways, so why not offload it from the CPU. For most video games, you are only simulating simple vector physics(ie what happens when I kick this box at x,y,z with N force and g gravity), so you don't need a radical new fancy expansion bus. Eventually you'd be able to simulate every significant object in a house when you slam a train into it.

ATI is claiming that x1900x owners can do this with a driver update.

So we have, in effect, two computers working in tandum with each other.

A good example of why offloading physics from the CPU would be a good idea would be the HL2 demo videos, the one with the barrels falling down the pegs. While just sending 5-9 barrels flying is easy enough for your CPU to calculate, crashing a train into them and having it look realistic requires more effort. First you have to run colision detection to figure out when and where the barrel was hit, using that figure out what force to apply to the barrel, was it hit flat on or at an angle, was it hit dead center or on an edge, all of those will tell you what way to send the barrel flying, what way to spin it as it flys off, etc. It's still within the bounds of what a CPU can do without too much work, you can even run AI over the top of it. But what happens if you turn it into a stockpile of 150 barrels? Or worse?

Now if you want to have some real fun, like flip's example, crash the same train into an ammo dump full of explosives and 500+ non static objects with dozens of static objects in the way of the now in flight barrels, crates, cars, trains and other debris while running the AI for 20-30 NPCs and playing 30-50 different 2d and 3d voices. For even more fun make it so the train, cars and barrels can be bent, crushed or torn apart, even taking all the shortcuts you can it is way more then a modern CPU can do in realtime.

Krom wrote:Now if you want to have some real fun, like flip's example, crash the same train into an ammo dump full of explosives and 500+ non static objects with dozens of static objects in the way of the now in flight barrels, crates, cars, trains and other debris while running the AI for 20-30 NPCs and playing 30-50 different 2d and 3d voices. For even more fun make it so the train, cars and barrels can be bent, crushed or torn apart, even taking all the shortcuts you can it is way more then a modern CPU can do in realtime.

mmm...

From what I understand of Physics processing, a CPU is not the ideal target processor to do this stuff: a CPU is far too generalised to be able to efficiently calc this sort of stuff on anything except a very small scale.

GPUs on the other hand are exceedingly specialised processors (Which, conversely would run Windows operating system code very slowly) which are especially designed to do vector math (among other things). Physics algorythms can leverage these very powerful math processors to provide a shozload more physics capability than the CPU is capable of.

Gotcha. Thanks for the explanation.

Krom wrote:A good example of why offloading physics from the CPU would be a good idea would be the HL2 demo videos, the one with the barrels falling down the pegs. While just sending 5-9 barrels flying is easy enough for your CPU to calculate, crashing a train into them and having it look realistic requires more effort. First you have to run colision detection to figure out when and where the barrel was hit, using that figure out what force to apply to the barrel, was it hit flat on or at an angle, was it hit dead center or on an edge, all of those will tell you what way to send the barrel flying, what way to spin it as it flys off, etc. It's still within the bounds of what a CPU can do without too much work, you can even run AI over the top of it. But what happens if you turn it into a stockpile of 150 barrels? Or worse?

Now if you want to have some real fun, like flip's example, crash the same train into an ammo dump full of explosives and 500+ non static objects with dozens of static objects in the way of the now in flight barrels, crates, cars, trains and other debris while running the AI for 20-30 NPCs and playing 30-50 different 2d and 3d voices. For even more fun make it so the train, cars and barrels can be bent, crushed or torn apart, even taking all the shortcuts you can it is way more then a modern CPU can do in realtime.

Yeah, you could process an equation for each object and be a dumbass like that ... or have one equation that handles all aspects in the game. There the bottleneck is memory and not so much processing.

If that were true ccp, crashing a wash machine into a pile of crates wouldn't cause the HL2 engine to freeze up for several frames. I even bet that for the most part the colision detection and physics processing that has to be done in HL2 all fits inside the L1 cache.