1. Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

Did AMD change the architecture of HD7000?

Discussion in 'AMD / ATI' started by Artas1984, Oct 18, 2012.

  1. Artas1984

    Joined:
    Feb 22, 2009
    Messages:
    211 (0.11/day)
    Thanks Received:
    94
    Logic states that 1024 shaders can not beat 1600 shaders clocked at same speeds in a card with the same amount of ROPS,TMUS, having the same memory bandwidth, right?.....:wtf:

    But no, i guess i am an idiot, because god help me to understand how on earth can HD7850 be faster than HD5870 with SIGNIFICALLY lesser shader power even on low resolutions like 1650:

    http://www.anandtech.com/show/5625/...-hd-7850-review-rounding-out-southern-islands

    I do not believe in things which do not make sense, so please try to explain in great detail the what is the secret behind this? :banghead:

    The increased transistor count adds raw FPS performance? :wtf:
    The die shrink adds raw FPS performance? :wtf:

    I mean come on, this is BS.
  2. TRWOV

    TRWOV

    Joined:
    Aug 11, 2011
    Messages:
    3,381 (3.17/day)
    Thanks Received:
    1,934
    Location:
    Mexico
    HD5000 series = VLIW5, the same architecture ATi had been using since R300.
    HD7000 series = GCN, new compute geared architecture. It has less shaders but stronger ones.
    Last edited: Oct 18, 2012
    Crunching for Team TPU
  3. Aquinus

    Aquinus Resident Wat-man

    Joined:
    Jan 28, 2012
    Messages:
    5,888 (6.56/day)
    Thanks Received:
    1,888
    Location:
    Concord, NH
    The process in which the GPU was manufactured is smaller on the 7k series cards which equates to lower power usage and higher speeds as well.
  4. W1zzard

    W1zzard Administrator Staff Member

    Joined:
    May 14, 2004
    Messages:
    14,632 (3.94/day)
    Thanks Received:
    11,363
    that

    smaller process does not make things faster (at same clock frequency), but might allow higher clock frequency
  5. qubit

    qubit Overclocked quantum bit

    Joined:
    Dec 6, 2007
    Messages:
    9,821 (4.07/day)
    Thanks Received:
    3,479
    The shaders in the later card are more advanced and can do more work per clock. Hence, a lower end model of a newer architecture can often beat the higher end card of an older architecture. For example, my GTX 285 only has about the same rendering power as a lower end GTX 460. On top of that, the 460 supports more features, such as DX11.

    It's exactly the same principle as not rating CPU's performance on clock speed alone, since the internal architecture makes a big difference.

    Therefore, it's not BS. :)
  6. Zubasa

    Zubasa

    Joined:
    Oct 1, 2006
    Messages:
    3,980 (1.40/day)
    Thanks Received:
    457
    Location:
    Hong Kong
    Even the 6950 preforms better than the 5870 with less shaders.
    The 5870 with its old vliw5 architecture have problems of under utilizing its shader power, and this is the one of the reasons why the 2900XT aka R500 failed.
    What AMD did with the 4870/RV770 was to simply add a lot more shaders to compensate.

    VLIW5 was good in theory but in games it simply have problems fully loading all of the shaders, the lower clocked 6950/Cayman with its simplified VLIW4 architecture is able to practically match/outperform the 5870 with less shaders because it is better utilized.
    The newer GCN bases HD7000 series have is much better at distributing the load to all of its shaders because in GCN the shaders are pretty much homogeneous.
    The older VLIW have complex heterogeneous shaders and have great theoretical shader power, but in reality we don't live in a perfect world. Games are just not coded to run specifically on AMD's hardware.

    Actually W1zzard have some good info in his 7970 review:
    http://www.techpowerup.com/reviews/AMD/HD_7970/2.html
    [​IMG]
    Last edited: Oct 19, 2012
    Artas1984 says thanks.
  7. eidairaman1

    eidairaman1

    Joined:
    Jul 2, 2007
    Messages:
    11,941 (4.65/day)
    Thanks Received:
    1,335
    7700-7900 i believe is GCN where 7600 and less are VLIW4
  8. TRWOV

    TRWOV

    Joined:
    Aug 11, 2011
    Messages:
    3,381 (3.17/day)
    Thanks Received:
    1,934
    Location:
    Mexico
    All HD7000 series are GCN. Only HD6930/50/70/90 were VLIW4. < HD6800 were VLIW5. Not sure about Turks though.
    Crunching for Team TPU
  9. mastrdrver

    mastrdrver

    Joined:
    Feb 24, 2009
    Messages:
    3,099 (1.58/day)
    Thanks Received:
    562
    That is wrong. All 7700 and up are GCN. The next APU from AMD (Kabini) will also have GCN shaders.
  10. eidairaman1

    eidairaman1

    Joined:
    Jul 2, 2007
    Messages:
    11,941 (4.65/day)
    Thanks Received:
    1,335
    i had already updated my stuff, see below
  11. Zubasa

    Zubasa

    Joined:
    Oct 1, 2006
    Messages:
    3,980 (1.40/day)
    Thanks Received:
    457
    Location:
    Hong Kong
    The HD7600 and below are just rebranded chips so they are the same as their HD6000 counterparts.
  12. cdawall where the hell are my stars

    Joined:
    Jul 23, 2006
    Messages:
    20,648 (7.09/day)
    Thanks Received:
    2,971
    Location:
    some AF base
    That has zero to do with anything...
  13. eidairaman1

    eidairaman1

    Joined:
    Jul 2, 2007
    Messages:
    11,941 (4.65/day)
    Thanks Received:
    1,335
    just means the amt of transistors that can be stuffed in the core and meet power and heat requirments
    HD64G says thanks.
  14. Derek12

    Joined:
    Jan 12, 2010
    Messages:
    1,096 (0.67/day)
    Thanks Received:
    163
    I though that smaller processes of the same die were faster or less latency as the circuit paths are smaller, is this wrong?
  15. Aquinus

    Aquinus Resident Wat-man

    Joined:
    Jan 28, 2012
    Messages:
    5,888 (6.56/day)
    Thanks Received:
    1,888
    Location:
    Concord, NH
    Yes and no.

    Smaller circuitry enables a GPU manufacturer to cram more transistors onto a single GPU. Also since the process is smaller, the distance from any one point to another will be shorter, so a signal will reach where it is going a little faster than it was before. Smaller circuitry tends to favor higher frequencies as well.

    All in all, a die shrink does more to help performance and power usage than it does to hurt it.
    HD64G says thanks.
  16. scorca New Member

    Joined:
    Aug 27, 2012
    Messages:
    16 (0.02/day)
    Thanks Received:
    0
    There is nothing to do with distance. An electron has a speed of light so it really doesnt matter if it 1 km or 1 nm. You wont notice any difference. But you were right about higher frequency. The basic principe of a transistor is a switch. The higher the switching frequency the higher is the output but also the more heat produced. But smaller size means less heat. By shrinking the size of the transistor, it allows higher switching frequency but less heat.
  17. qubit

    qubit Overclocked quantum bit

    Joined:
    Dec 6, 2007
    Messages:
    9,821 (4.07/day)
    Thanks Received:
    3,479
    At ghz frequencies, those tiny distances do make all the difference. That's why smaller circuits clock higher.
  18. de.das.dude

    de.das.dude Pro Indian Modder

    Joined:
    Jun 13, 2010
    Messages:
    7,476 (5.01/day)
    Thanks Received:
    1,936
    wrong. electrons are no where near the speed of light.

    your concept is all messed up. i would correct it. but its past 1am and i am sleep.y

    instead heres a wikilink
    http://en.wikipedia.org/wiki/Drift_velocity

    and a google result
    https://www.google.co.in/search?cli...rceid=opera&ie=utf-8&oe=utf-8&channel=suggest


    this is highschool physics. their motion is actually hapazard and not in any particular constant direction. thats why the effective speed is slow.


    but when you switch on a light, it doesnt mean that an electron from the battery flows all the way to the light. in fact what happens is one electron knocks out the other. and there is a chain like this and this is how the energy gets transmitted so quickly.


    its true you can make the electron go at the speed of light but that depends on how much energy you give them. and 12V isnt a lot of energy :p
    Last edited: Oct 19, 2012
    qubit says thanks.
  19. AlienIsGOD

    AlienIsGOD

    Joined:
    Aug 9, 2008
    Messages:
    3,411 (1.58/day)
    Thanks Received:
    1,456
    but not the same VLIW 5 as the rest of the 6K series. Barts shaders are more refined over 5800 series. 5830 =1120 shaders, 6870 = 1120 refined VLIW 5, which explains the performance difference between 2 cards with the exact same number of shaders
  20. scorca New Member

    Joined:
    Aug 27, 2012
    Messages:
    16 (0.02/day)
    Thanks Received:
    0
    But it's not gonne be any faster (in speed not frequency)if you shorten the length.
  21. scorca New Member

    Joined:
    Aug 27, 2012
    Messages:
    16 (0.02/day)
    Thanks Received:
    0
    Ok, I put it wrong. I should say "energy or signal travelling in a conductor" instead of electron.
  22. qubit

    qubit Overclocked quantum bit

    Joined:
    Dec 6, 2007
    Messages:
    9,821 (4.07/day)
    Thanks Received:
    3,479
    Yes it will if you increase the clock frequency. Have you not understood DDD's detailed explanation?

    If you don't increase the clock frequency, then the chip will have a slightly lower latency but run at the same speed, because all the clock signals are gated. However, if it was a clockless design, like some ARM chips a few years ago, then it will indeed automatically run faster as you shrink it.
  23. scorca New Member

    Joined:
    Aug 27, 2012
    Messages:
    16 (0.02/day)
    Thanks Received:
    0
    I think we are talking about different thing here. Speed is not equal to frequency.
    Last edited: Oct 20, 2012
  24. Phusius

    Phusius New Member

    Joined:
    Mar 23, 2012
    Messages:
    1,318 (1.57/day)
    Thanks Received:
    380
    this topic makes me want to get laid. too much nerdfest.
    eidairaman1 and qubit say thanks.
  25. qubit

    qubit Overclocked quantum bit

    Joined:
    Dec 6, 2007
    Messages:
    9,821 (4.07/day)
    Thanks Received:
    3,479
    Ok, explain yourself! :toast:

    What do you mean then? :)

Currently Active Users Viewing This Thread: 1 (0 members and 1 guest)

Share This Page