Physics Processor - AGEIA (Cool Demo)

Discussion in 'GPGPU Technology & Programming' started by rwolf, Mar 8, 2005.

  1. CMAN

    Regular

    Joined:
    Mar 23, 2004
    Messages:
    620
    Likes Received:
    15
    Location:
    St Louey
    The building explode entrhalled me. I watched it a few times. I'm not crazy though, I wouldn't want to see that stuff in real life.
     
  2. digitalwanderer

    digitalwanderer Dangerously Mirthful
    Legend

    Joined:
    Feb 19, 2002
    Messages:
    17,396
    Likes Received:
    1,862
    Location:
    Winfield, IN USA
    If I could pick up a $100-200us priced card that could add super-cool physics to my games I think I'd do it in a heartbeat. :)
     
  3. Jawed

    Legend

    Joined:
    Oct 2, 2004
    Messages:
    10,873
    Likes Received:
    767
    Location:
    London
    There's a cute bug in the Golem ragdoll.

    If you grab it by the top of a foot (maybe other places...?) and swing it round a bit, you'll quickly find there's, ahem, perpetual motion...

    Right-click to grab.

    Jawed
     
  4. Richard

    Richard Mord's imaginary friend
    Veteran

    Joined:
    Jan 22, 2004
    Messages:
    3,508
    Likes Received:
    40
    Location:
    PT, EU
    Take it easy, I think you're reading way too much into my one liner. :roll: I was taken aback by this figure. Doesn't mean I'm passing judgement on whether or not it's worth it. For one, I'd like to see which games will make use of it and what performance/feature difference the PPU will provide.
    Secondly, the video card analogy is not that good because without one you can't play any games; without a PPU you'll still be able to play them (at least until/if PPUs become a requirement like 3D acceleration today) so I need to know if such an investment is sound by the conditions of the market. That's what I always try to do: don't reject something out of hand nor accept it blindly.

    So, there was no innuendo in my previous post so please don't "go look for reasons" there yourself. :wink:
     
  5. CMAN

    Regular

    Joined:
    Mar 23, 2004
    Messages:
    620
    Likes Received:
    15
    Location:
    St Louey
    Sorry, no harm meant. 8)

    I would want to see reviews for it before I bought it of course, but I'm excited for what it COULD do.
     
  6. RejZoR

    Regular

    Joined:
    May 9, 2004
    Messages:
    300
    Likes Received:
    3
    Location:
    Europe\Slovenia\Ljubljana
    Also try Physics submenu demo (access it by F4 button).

    When you're there press "M" key for like 10 seconds (or for as long as your CPU allows).
    I also recommend to first create Castle environment (Create Castle) for even better collision presentation.

    Check this out :)
    [​IMG]

    [​IMG]
     
  7. Dr. Ffreeze

    Regular

    Joined:
    Jun 6, 2003
    Messages:
    335
    Likes Received:
    0
    PPU Concerns,

    Three levels of cards, entry, mid, and high end. I would guess that they are discussing the specs of the high end model so a very from the hip conservative guess might make the cards look like this:

    High End: $400 400MHz 125M Transistors 40,000 entities
    Mid: $200 200MHz 125M Transistors 20,000 entities
    Entry: $100 100MHz 125M Transistors 10,000 entities
    CPU: $420 3400MHz 169M Transistors 40 entities

    How did I come up with these numbers, well I mostly guessed but here is why I guessed the way I did.

    High End:
    - $400: they said so at tomshardware
    - 125M transistors: they said so
    - 40,000 entities: I think they would be doing PR with their best part. and they were using 40,000 entities in their PR blitz.
    - 400MHz: Guess. This is between NVIDIA's (400MHz) and ATI's (520MHz) high end GPU speed. As TSMC is an investor it would be in there best interest to get this working well.

    Mid:
    - $200: Guess. Good midrange price point (for video cards anyway)
    - 125M transistors: Guess. I would guess it would make the most since for them to speed bin their parts instead of having different cores. This way they could supply the few fast parts to the high end while the bulk would be for the mid and entry cards.
    - 20,000 entities: Guess. Worst case scenario would be ~1/2 the speed of the high end part. While it is much lower than their PR it still kicks a high end CPU's butt (if their PR is to be believed).
    - 200MHz: Guess. If their high end parts could hit 400MHz, then the yields would be much better at 1/2 speed, plus it makes the numbers nice. =)

    Entry:
    - $100: They said so in toms hardware that ~$100 could be the low end range of prices.
    - 125M transistors: Guess. I would guess it would make the most since for them to speed bin their parts instead of having different cores. This way they could supply the few fast parts to the high end while the many would quall for the mid and they could salvage the worst performers for the entry level.
    - 10,000 entities: Guess. Worst case scenario would be ~1/2 the speed of the mid ranged part. While it is much, much lower than their PR it still kicks a high end CPU's butt (if their PR is to be believed).
    - 100MHz: Guess. If their high end parts could hit 400MHz, then the yields would be much better at 1/4 speed, plus it makes the numbers nice. =)

    CPU:
    - $420: Pentium 4 650 at www.newegg.com
    - 169M transistors: so says techreport.com
    - 40 entities: so says Ageia's PR. Seems a bit low but I guess it depends on the game. I would guess HL2 would have more than 40 entities you could interact with.
    - 3,400MHz: So says Intel

    Ok, bases on those seat of my pants numbers I have few questions/thoughts.

    Take 1: Lets say the Unreal 3 Warfare engine (which will support the new PPU) will only use the add PPU to increase eye candy. Lets Epic intros a new flame thrower that is plain amazing. The high end card renders it with 40,000 entities, 20,000 for the mid, 10,000 for the entry, and the lowly CPU gets the age old plane texture flame thrower. Well, regardless how your level of physics complexity all four types of players could play online together well. The problem I see here is that it is does not make a very good case to buy the $400 Ultra PPU.

    Take 2: Epic decides to make a game that shows off how utterly cool it would be to have 1000 times the entities you can interact with and this actually changes how you play the game, how you solve puzzles. This might allow you to make the assault from the back door, when there was never a door in the back to begin with. You just blow your own door. You could instead make it a side assault or even one from the roof. This would be a sweet reason to buy the PPU Ultra card but how is someone with a lowly 3.4GHz CPU going to be able to play the game?

    If we don't see some very compelling reasons to upgrade, more reasons that just eye candy, then I believe Ageia will have a very hard time succeeding. The problem is, I don't see how game programmers will be able to balance the game experience between those that have a PPU and those that do not if they use the PPU power for more than just eye candy.

    Thoughts?
     
  8. Dr. Ffreeze

    Regular

    Joined:
    Jun 6, 2003
    Messages:
    335
    Likes Received:
    0
    Chalnoth,

    I could would, could you at least read the last bit of my last post? I believe you are talking about what I am also concerned with. Unless it is only eye-candy how will they balance between the PPU haves and have nots?
     
  9. Jawed

    Legend

    Joined:
    Oct 2, 2004
    Messages:
    10,873
    Likes Received:
    767
    Location:
    London
    The implications are pretty clear here. You have either 30fps (or less) without PPU or 100fps (or more) with one.

    As games get more complex, here's a way to scale up gaming performance, independently of screen resolution, or how many cores your CPU has. Dual-core, when it comes is eventually going to be an approximate doubling of gaming performance in CPU-hogging physics (well, less, due to other aspects of a game such as AI).

    This is a great way to un-tie another dimension of game programming from the PC's CPU-performance brickwall.

    Games aren't just graphics. There's also no reason why gaming consoles in 3-5 years' time wouldn't use such a PPU component in their architecture.

    Jawed
     
  10. Richard

    Richard Mord's imaginary friend
    Veteran

    Joined:
    Jan 22, 2004
    Messages:
    3,508
    Likes Received:
    40
    Location:
    PT, EU
    I think your core speeds are too apart. I can certainly see the low-end having only 64mb onboard ram or whatever and have nearly the same core speed as that mid-level part. Also, since Aegis will only supplies the chipsets themselves we could potentially see a lot of different designs from all the AIB partners. Which reminds me, what makes you guess that the low-end parts will have the exact same chip (125m tranny count)? It makes sense for the startup but I think they'll want to ramp out different designs for the different price tiers as soon as they can.

    Since they talked about mobo integrated PPUs I'm positive they have different designs already.
     
  11. Dr. Ffreeze

    Regular

    Joined:
    Jun 6, 2003
    Messages:
    335
    Likes Received:
    0
    Mordenkainen,

    I agree, I just wanted to go with the "worst case scenario". It could be argued that the MHz doesn't matter really, just the end result in performance. They could fill the entry level boards with a faster part but with just less memory bandwidth and/or amount like you said.

    The only thing would be the ease of pulling this off at the expense of a more costly entry level chip. They could also have a 50m tranny count chip for the entry level and save money on each chip that went out the door, it would just be more complicated to design 2 different chips for a startup (which they could do, my idea was just a guess).

    That is a very good point that I had not thought of.

    I also find it interesting if they could release a $50 board. If they could do this and it had a very low performance compared to their other offerings but still had a 2 orders of magnitude greater performance than CPUs, they would be able to move many many chips. At $50 it is the same price as a quality new release game. I don't think they would be able to do it, but it would be nice for their install base.
     
  12. KimB

    Legend

    Joined:
    May 28, 2002
    Messages:
    12,928
    Likes Received:
    230
    Location:
    Seattle, WA
    Because with graphics you could change quality levels to coincide with your hardware. With physics there's just no good way to change quality smoothly in the same way.
     
  13. KimB

    Legend

    Joined:
    May 28, 2002
    Messages:
    12,928
    Likes Received:
    230
    Location:
    Seattle, WA
    Has nothing to do with the API. The API may be a good thing (as it could possibly lead to better usage of current processors, CPU or GPU, for physics), but dedicated hardware just doesn't make any sense for the PC. The reason is simple: physics is either on or it's not (for the most part). There's no good way to have different quality levels. Therefore developers must program for the lowest common denominator, which will always mean PC's without physics accelerating hardware.
     
  14. DeanoC

    DeanoC Trust me, I'm a renderer person!
    Veteran Subscriber

    Joined:
    Feb 6, 2003
    Messages:
    1,469
    Likes Received:
    185
    Location:
    Viking lands
    Thats assuming physics are a bottle-neck, we find that in a complex scene physics (Havok 2.3) took up 5-10% CPU time, thats everything collision, response, ragdolls and constraints.

    So even if the PPU can remove that completely, we will have gone from 30fps to 33fps...

    To do more physics, we have a little problem, the problem is that we have to render those individually physic'ed objects, so you also have to speed up rendering. So if you want better physics, first you need to reduce the cost of rendering again (ironic how all the its not all rendering statements end up meaning its all about the rendering ;-) )...

    Makes no difference if there system can do 40,000 objects per frame, we could never render them... Also its worth noting most good physics engines can do 100s if not 1000s right now, its the number of interacting (colliding) entities that slow things do (20 interacting ragdolls (body heap) of ~20 bones each is about the normal CPU limit).

    Rock meet hard place...

    I can see the PC maybe needing a physics chip (if we can't get the GPU taking some of the burden, I suspect they will though...), as your only getting 2 cores near term. Don't see the need on the consoles (except maybe Revolution... god knows what thats doing). Its not like we are short of processing power (using it effectively is a harder problem of course) on them.

    And while the publicity has been on Aegis/Novodex and Meqon this GDC, I just want to say how good the new Havok 3 is. We are currently doing a quick intergration with it and continous collisions rock. Sure they won't mind me saying that ;-)

    Other interesting thing, is if this PPU is doing discrete or continous collision. I hate discrete collisions system with a passion but most physics engines use them (a lot harder to generate contact points)

    Sorry rambling a bit... Just a general musing rather than any well thought out comments.
     
  15. Jawed

    Legend

    Joined:
    Oct 2, 2004
    Messages:
    10,873
    Likes Received:
    767
    Location:
    London
    In Novodex Rocket, with the Building Explode demo, running on A64 3500+ with 9800Pro 128MB, 256-bit (6000-ish 3DMk03, "unoptimised") I get the following frame rates in 1024x768 windowed mode, at about 10 seconds in:

    - 12fps with the physics simulation running

    - 100fps with the physics simulation paused

    - 50fps with the physics simulation paused, 4xAA

    Clearly my graphics card is running well below capacity simply because the CPU is struggling with the physics.

    This is an extreme example, of course (4200 bricks - number of bricks varies, not sure why...). But the CPU isn't doing much else...

    How game developers might use, say, 1-2000 "fragments" of a solid object being blown apart, or 1-200 bodies (say) being thrown around I don't know...

    Jawed
     
  16. KimB

    Legend

    Joined:
    May 28, 2002
    Messages:
    12,928
    Likes Received:
    230
    Location:
    Seattle, WA
    It is always possible, of course, that that demo's physics engine isn't as efficient as the one from Havok.
     
  17. ERP

    ERP Moderator
    Moderator Veteran

    Joined:
    Feb 11, 2002
    Messages:
    3,669
    Likes Received:
    49
    Location:
    Redmond, WA
    Your kind of missing Deano's point.

    Outside of trivial demos there are constraints besides "time it takes to compute physics" in a real game. On a PC in particular drawing a lot of objects is an extremly expensive operation, 40,000 objects means probably >40,000 calls to DrawPrimitive, even with a NULL driver this is extremly expensive.

    If your goal is enabling more moving objects, you have to solve a lot more than just the physics cost.
     
  18. Killer-Kris

    Regular

    Joined:
    May 20, 2003
    Messages:
    540
    Likes Received:
    4
    That depends on what you're doing with your physics. If you're doing inverse-kinematics for animations that interact with the environment (stairs, uneven ground, etc), you just use a lower resolution bone structure for the lower performance PPUs, or disable the inverse-kinematics for CPU users.

    Same goes for any fluid simulations (bodies of water, clouds, etc...), you use a lower resolution of points on the lower end part, and just have a static fluid surfaces for CPU users.

    And I'm sure that there are more creative solutions to scale physics performance than what I've listed. Of course there may not be any sort of good solution to the deformable terrain/buildings to handle the CPU users. So until PPUs become ubiquitous they probably will only serve to make a prettier and more immersive environments. But I imagine that many of these effects will have the same effect on people’s opinions, that bilinear filtering had when it was new. So with any degree of luck the user base will build quickly.
     
  19. rwolf

    rwolf Rock Star
    Regular

    Joined:
    Oct 25, 2002
    Messages:
    968
    Likes Received:
    54
    Location:
    Canada
    That is a good point. Not only that but you must ship 40,000 objects from the physics card to the video card for each frame. At some point the CPU has to manipulate these objects as well.
     
  20. rwolf

    rwolf Rock Star
    Regular

    Joined:
    Oct 25, 2002
    Messages:
    968
    Likes Received:
    54
    Location:
    Canada
    Who knows the answer to that?

    However this API should rock if you have hyper threading or dual cpus.
     
Loading...

Share This Page

  • About Us

    Beyond3D has been around for over a decade and prides itself on being the best place on the web for in-depth, technically-driven discussion and analysis of 3D graphics hardware. If you love pixels and transistors, you've come to the right place!

    Beyond3D is proudly published by GPU Tools Ltd.
Loading...