HD 4870 review thread.

Discussion in '3D Hardware, Software & Output Devices' started by mczak, Jun 24, 2008.

  1. sireric

    Regular

    Joined:
    Jul 26, 2002
    Messages:
    348
    Likes Received:
    22
    Location:
    Santa Clara, CA
    Well, for this specific game, I have no idea "why" :)-)). I was just trying to illustrate a case where two GPUs could outperform by more than 2x a single GPU. Yes, it would require them to be working on the same frame in the simplest view, so that a "given" texel is only required by 1 GPU, so that the caches are effectively larger (due to the reduced load).

    For dual GPUs in AFR, it does seem harder to come up with real cases for super-linear. There's some cases, like during transitions, that each GPU has unique data in their local memory, which can lead to higher total local memory for the pair than each individual (usually, the dual GPU are done to get 2x the BW). That can give super-linear, though it seems very specific. I'll have to think about this some more.
     
  2. mczak

    Veteran

    Joined:
    Oct 24, 2002
    Messages:
    3,015
    Likes Received:
    112
    Too bad :).
    This is all well and good but it does not solve the fundamental issue of output lag (e.g. two gpus running at 30fps have the same output latency as a single one running at 15fps would have). Doesn't matter for non-interactive work or if your framerate is very high (but then you wouldn't really need two gpus to begin with...), of course.
     
  3. sireric

    Regular

    Joined:
    Jul 26, 2002
    Messages:
    348
    Likes Received:
    22
    Location:
    Santa Clara, CA
    I agree. Though, in general, for 2 GPUs this isn't too bad (subjective, of course).
     
  4. MfA

    MfA
    Legend

    Joined:
    Feb 6, 2002
    Messages:
    6,811
    Likes Received:
    478
    Okay so that pretty much rules out a really high bandwidth crossfire link then ... ah well, can't have everything :) (Still stings a little though.) Perhaps the last guess on the closed architecture thread was correct and it's an extra PCI-E connection.
     
  5. Mintmaster

    Veteran

    Joined:
    Mar 31, 2002
    Messages:
    3,897
    Likes Received:
    87
    I'm pretty sure this one case is due to error and/or platform differences. The single GPU is Nforce, and Crossfire is done on X48 or something.

    I still remember articles harping about latency in the Rage Fury MAXX days. :smile: NV's SLI got a free pass, though, for a long time.

    Do you have methods to align the frame output evenly? I still hear people mentioning stuttering issues, i.e. pairs of frames hitting the display too close together.
     
  6. Mintmaster

    Veteran

    Joined:
    Mar 31, 2002
    Messages:
    3,897
    Likes Received:
    87
    I'm hoping that in a few generations we'll see terabit optical links between chips. :wink:
     
  7. DemoCoder

    Veteran

    Joined:
    Feb 9, 2002
    Messages:
    4,733
    Likes Received:
    81
    Location:
    California
    It is not a graphic card feature, PhysX is a middleware library, originally the Novodex Physics Engine that ran purely via CPU. You code to the library, and if you have a PS3, it runs on CELL accelerated, if you have a multicore CPU, it leverages the extra CPU cores to run physics, if you have a PPU, it runs there, and now if you have an NVidia GPU, it runs accelerated on the GPU. If you have an AMD card, you are not left out, you just have to settle for CPU acceleration.

    There are tons of defacto-standard middleware libraries that are licensed and OEMed by game-devs. There don't have to be "standardized" to become very widespread. Look at the omnipresent Bink and MILES libraries. Havok seems included in about 50% of titles now.

    Intel could very well release a "Larrabee Havok" that accelerates Havok alot faster if and only if you have an Intel CPU. You think this won't catch on? I don't. Sure, it will eventually be replaced by something like OpenCL or compute-shader implementations, but those are very far out, and Developers have a need today to solve these problems.

    PhysX supports both CPUs and GPUs as well. Which one wins will come down to which has the better API and implementation. Havok has marketshare today, but if CUDA is far better at Physics than CAL is, developers will be hard pressed to ignore CUDA, especially since CUDA is far more mature.
     
  8. Megadrive1988

    Veteran

    Joined:
    May 30, 2002
    Messages:
    4,638
    Likes Received:
    148
    Two 4850X2 / 4 TeraFLOPs - $999 ~ $1099
    Two 4870X2 / 5 TeraFLOPs - $1199 ~ $1249

    Those would be nice options.
     
  9. Mintmaster

    Veteran

    Joined:
    Mar 31, 2002
    Messages:
    3,897
    Likes Received:
    87
    Havok's marketshare means that PhysX must be a lot faster to win, particularly because physics speed is not really a limiting factor except for gimic effects. Having good tools, documentation, support, etc. is the reason Havok is so dominant today, because otherwise people would be choosing cheaper/free solutions.

    Besides, I see little chance that Havok won't be as fast or faster in speed. They've been in the game a lot longer than Ageia, and solving physics quickly is largely about incorporating little tricks and initial guesses for the particular situations seen in games.

    The only way I see PhysX taking off is if Intel pulls the plug on GPU support. I don't see that happening for a long time, because they need Havok to remain dominant if they want Larabee to have a selling point in physics. Of course, we never know...
     
  10. jimmyjames123

    Regular

    Joined:
    Apr 14, 2004
    Messages:
    810
    Likes Received:
    3
    How's this for a possible Radeon lineup at 55nm sprouting from just a single RV770 GPU:

    $1199 CrossfireX 4890 2048 MB GDDR5
    $899 CrossfireX 4880 2048 MB GDDR3
    $599 4890 1024 MB GDDR5 (= "4870 X2")
    $449 4880 1024 MB GDDR3 (= "4850 X2")
    $299 4870 512 MB GDDR5
    $199 4850 512 MB GDDR3
     
  11. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    8,186
    Likes Received:
    1,841
    Location:
    Finland
    I'm quite sure the lineup will be like this:
    $199 4850 512MB GDDR3
    $229-249 4850 1GB GDDR3
    $299 4870 512MB GDDR5
    $329-359 4870 1GB GDDR5
    $499-549 4870X2 1GB GDDR5
    $549-599 4870X2 2GB GDDR5

    If we'll see 4850X2, it will probably be AIB specific solution like with 38x0's
     
  12. Jawed

    Legend

    Joined:
    Oct 2, 2004
    Messages:
    10,873
    Likes Received:
    767
    Location:
    London
    Physically, as per the die shot, it looks completely different.

    Jawed
     
  13. sireric

    Regular

    Joined:
    Jul 26, 2002
    Messages:
    348
    Likes Received:
    22
    Location:
    Santa Clara, CA
    Oops, I take that one back. Quakeworld does scale well in general, but a case was found where it doesn't (illustrated above) and a fix should be available in an upcoming driver.
     
  14. DemoCoder

    Veteran

    Joined:
    Feb 9, 2002
    Messages:
    4,733
    Likes Received:
    81
    Location:
    California
    I don't think it's that simple, due to the complicated relationship between IHVs, middleware ISVs, and engine ISVs. For example, if NVidia pays Epic for first-class PhysX support in Unreal, and Unreal Engine "wins" in the sense that the most popular triple-A titles license it, then NVidia "wins" in the sense that most of the games being benchmarked will have some PhysX support.

    There is also the possibility that NVidia will simply eliminate licensing fees and open source PhysX at some point. From a software heritage point of view, Intel is alot worse at developer support than NVidia, so it's possible PhysX will gain ground.


    My point is, a "glide" strategy can succeed as long as there is chaos in the market. Glide worked for years until the 3dfx minigl driver came about and helped get DX and OGL to displace it. NVidia's Cg was only defeated because GLSL and HLSL got adopted as standards (and relatively quickly). If OpenCL/Khronos Equivalent/DirectX11 compute shaders don't arrive quickly enough, there will be a vacuum left in the market for vertical solutions, like PhysX, Elemental's transcoder, and other CUDA leveraged apps. When do you expect DX11 compute shaders that can compete with CUDA? 2009? 2010?

    I think there is the very real possibility that you will see triple-A TWIMTBP engine licensees incoporating PhysX middleware prior to Havok GPU acceleration hits the market for AMD.
     
  15. waver_01

    Newcomer

    Joined:
    Jun 24, 2008
    Messages:
    5
    Likes Received:
    0
  16. AlphaWolf

    AlphaWolf Specious Misanthrope
    Legend

    Joined:
    May 28, 2003
    Messages:
    8,486
    Likes Received:
    334
    Location:
    Treading Water
    just saw this

    looks like idle power use and idle heat will be under control.

    <edit> although that seems to be a 4850, and overclocked.
     
    #96 AlphaWolf, Jun 26, 2008
    Last edited by a moderator: Jun 26, 2008
  17. XMAN26

    Banned

    Joined:
    Feb 17, 2003
    Messages:
    702
    Likes Received:
    1
  18. Dave Baumann

    Dave Baumann Gamerscore Wh...
    Moderator Legend

    Joined:
    Jan 29, 2002
    Messages:
    14,079
    Likes Received:
    648
    Location:
    O Canada!
    I'm not sure you've been following what Havok have been doing recently, but if you have its pretty obvious where things will go (on both Intel and AMD GPU).

    Additionally, Havok's core middleware has been opened up license free to PC devs.
     
  19. Geo

    Geo Mostly Harmless
    Legend

    Joined:
    Apr 22, 2002
    Messages:
    9,116
    Likes Received:
    213
    Location:
    Uffda-land
    I think NV's devrel record is pretty impressive. Even if it takes another two years before it starts to pay off with much visibility. Which doesn't necessarily imply dominance, but I think their track record suggests it is likely they'll gain a substantial foothold for Physx over time.

    Edit: Roy Taylor:

    Does Taylor have an interest in "bigging up" the ISV numbers? Well, duh. But there's usually a grain of truth in these things as well.
     
  20. apoppin

    Regular

    Joined:
    Feb 12, 2006
    Messages:
    255
    Likes Received:
    0
    Location:
    Hi Desert SoCal
    i am buying *two* - am i NUTS?

    i am late here

    But color me impressed with AMD. The 4870 has killed any idea of me getting a GT260 and i would much rather get a 2nd 4870 for Crossfire than a Single GTX280 for the same price.

    i don't think Tesla is going to make Nvidia a lot of money this round and i LOVE the new resurrected competition that began with 38X0 and continues much improved with this series

    Of course, i am looking forward to the new "b" 280 against the 4870x2 [which i expect ASAP from AMD; as soon as they can deliver it]

    http://www.newegg.com/Product/Product.aspx?Item=N82E16814102748

    $309 plus shipping; $340.05 with tax and shipping total!

    Talk me out of it .. i have *two* in my basket; shall i get just one?
    - i only have the slower MB with 1 16x PCIe slot and the slow one. One now and a 4870x2 when i upgrade?

    Final edit:

    i am getting just ONE .. my sense has returned. It IS the equal of the GTX260 and i want a brand new GPU - just "because" [i am celebrating; i got my insurance settlement fully in my bank account and paid my high interest credit to zero]
     
    #100 apoppin, Jun 26, 2008
    Last edited by a moderator: Jun 26, 2008
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...