Myth debunking (or bunking!) thread. Fermi/kepler/maxwell vs gcn driver perf thread

Discussion in 'Benchmark Mayhem' started by Ieldra, Apr 13, 2016.

  1. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    POST UPDATE 2

    AS FAR AS OVERCLOCKING GOES:

    If you want to post overclocked results feel free to do so, HOWEVER, you ***MUST*** post stock results.

    You will be committing to A BENCHMARK not to a driver. If you say you will bench FFXIV HEAVENSWARD you'll bench it over a series of drivers!
    As far as driver settings go, EVERYTHING AT DEFAULT.

    This means:
    No IQ settings being changed in nvidia/amd control settings
    No tesselation sliders

    Everything at DEFAULT.


    Technically anyone can participate.

    Even nVidia Fermi/Maxwell owners. (If Maxwell owner knows how to add his card to 'unsupported' older driver.) And Fermi owner will add few more Driver versions above what will be official 'test set'. Namely starts with release driver for given card (if applicable for tested OS in any way), performance boost driver somewhere along the way (if any).
    Same applies for AMD VLIW4/5 owners.



    And word of advice, Keep your GPU at stock clock, it is about ratio/improvement between drivers. Not about people burning GPU due to old driver which accidentally forgot to spin fan over 20% of speed under load. (Use Afterburner with custom Fan curve at all times for your safety.)

    Official test set for NVIDIA cards.
    Official test set for AMD cards.
    Links for 13.3b3/12.11b11 have '*****' which corresponds to 'ngo hq' without space in middle.
    - 2nd link is provided because local g3d link appeared to be broken.
    - only allowed modification is to add your card into supported set (to allow driver installation)

    Official set of Benchmarks (high availability).
    Get banchmarks, ready. Do not benchmark blindly without speadsheet to fill. (UPDATE3)
    Document has several sheets. Benchmarking participants are to fill only cells with yellow background. 'TestSetup' is one where you fill your info.
    Each benchmark has its own sheet.
    - And has 2 sets for different CPU clock (if you are into comparing impact of CPU as well). ThrottleStop can disable Turbo on intel's CPUs without need to reset system.
    - Cell A1 contains benchmark API to test. Cell B1 has link in case you need it.
    - Benchmarks have pre-filled 'Driver Package Version' and results as an example. Please replace those with your values.
    - Sheet 'Drivers' has set of links for driver download (just in case)

    IMPORTANT:

    - Once you decide on settings for given benchmark (resolution, details). Do not change it in older drivers even if it means bad performance.
    - We are testing to see differences drivers make on same hardware.

    UPDATE 3: OpenOffice spreadsheets
    OneDrive Shared Folder.
    - nVidia drivers are still not selected. Once I get those missing driver versions/links, I'll get you nVidia version (please, do not waste your time on random drivers).[/QUOTE]
     
    Last edited: Apr 16, 2016
  2. -Tj-

    -Tj- Ancient Guru

    Messages:
    18,097
    Likes Received:
    2,603
    GPU:
    3080TI iChill Black
    I benchmarked few when I first got 980TI ~ 2months ago..

    But personally I didnt see any real driver regression, always checked with futuremark API benchmark, ok the fastest driver api overhead driver was definitely 353.52? @ 780GTX (dx11 final score), but by actual drawcalls it didnt change since r350+, even now with 980Ti @ dx11 api part..


    Tests
    OK, I've done some testing compared to my old OC'ed GTX780 to max vs stock 980ti boost 1342mhz

    Mostly lots of cpu bound benchmarks


    FFXIV - heaven shard

    GTX 780 1228mhz
    [​IMG]

    980Ti
    [​IMG]


    Hitman Absolution, ultra & 2xMSAA,.. CPU bound here now

    GTX 780 1215mhz
    [​IMG]

    980TI
    [​IMG]



    MHO Online

    GTX780 1215mhz
    [​IMG]

    980TI
    [​IMG]


    RE6

    GTX780 1215mhz
    [​IMG]

    980TI
    [​IMG]


    3dmark Skydiver

    GTX780 1228mhz, cpu 4.7ghz

    [​IMG]

    980TI, cpu score was a bit lower, maybe its a win10 thing, well that cpu score by 780 was best of (got a little lower too time to time)
    [​IMG]



    TR2013, ultimate @ FXAA
    GTX 780 1215mhz
    [​IMG]

    980TI
    [​IMG]


    Vally @ Extreme HD

    GTX 780 1215mhz
    [​IMG]

    980TI, funny how it reads 1530mhz boost, internal max boost bios setting, real stock 1342mhz
    [​IMG]
     
    Last edited: Apr 13, 2016
  3. Fox2232

    Fox2232 Guest

    Messages:
    11,808
    Likes Received:
    3,371
    GPU:
    6900XT+AW@240Hz
    Technically anyone can participate.
    Even nVidia Fermi/Maxwell owners. (If Maxwell owner knows how to add his card to 'unsupported' older driver.) And Fermi owner will add few more Driver versions above what will be official 'test set'. Namely starts with release driver for given card (if applicable for tested OS in any way), performance boost driver somewhere along the way (if any).
    Same applies for AMD VLIW4/5 owners.

    @-Tj-: May you please throw that huge quote into [spoiler ] tag?
    And since you are here, can you propose set of nVidia drivers to test? (It is not AMD vs nVidia thread. It is nVidia vs nVidia and AMD vs AMD. So it is not about picking best there is, but picking important drivers.)
    Basically idea is like this:
    16.4.1 – Crim - Newest
    15.11 – Crim - 1st
    15.11.x – ccc - last ccc
    15.7 – ccc – Win10 driver (wddm 2.0)
    14.12 – ccc – Omega
    13.3b3 – ccc – HD7790 release
    12.11b11 – final performance driver from series
    12.2 – Stable GCN
    11.12 – release GCN
    That's 9 drivers and it is far more than enough or desirable. Since switch from CCC to Crimson kind of requires comparison checkpoint AMD needs 2 more in list, WDDM 2.0 driver is important, Omega is supposedly important and release of that GCN1.05beta hardware in HD7790 (mostly buggy cards) had some changes too - probably visible only on GCN 1.1 and newer.
    Basically I think nVidia may need like ~6 drivers tested from Kepler launch till today. Including milestone like Maxwell 1.0/2.0 launch driver (to make testing easier for those owners).

    Original benchmarks proposition list:
    CS:Source - pure CPU bound and probably least valuable due to age
    AvP2010
    Metro 2033 - CPU bound if benchamarked with PhysX CPU
    WH40k: DoW2 - Retribution
    UnrealEngine 3 China Streets Benchmark - replaced Arma 2 (Standalone)
    Final Fantasy XIV - World Benchmark (Standalone)
    Resident Evil 6 Benchmark (Standalone)
    Tomb Raider
    Monster Hunter Benchmark (Standalone)
    Final Fantasy XIV - Heavensward (Standalone)
    Catzilla Benchmark (Standalone)
    5 Games
    5 Free Game Benchmarks
    1 'Free' benchmark
    (Would love to add one more from 2016, optimally Free to get as paid may reduce number of available testers drastically.)

    And word of advice, Keep your GPU at stock clock, it is about ratio/improvement between drivers. Not about people burning GPU due to old driver which accidentally forgot to spin fan over 20% of speed under load. (Use Afterburner with custom Fan curve at all times for your safety.)

    UPDATE: Official test set for AMD cards.
    Code:
    16.4.1 – Crimson:
    http://www.guru3d.com/files-details/amd-radeon-software-crimson-16-4-1-driver-download.html
    
    15.11 – Crimson:
    http://www.guru3d.com/files-details/amd-radeon-software-crimson-edition-download.html
    
    15.11.x – ccc:
    http://www.guru3d.com/files-details/amd-catalyst-15-11-1-beta-driver-download.html
    
    15.7.1 – ccc – Win10 driver:
    http://www.guru3d.com/files-details/amd-catalyst-15-7-driver-download.html
    http://www.techspot.com/downloads/drivers/essentials/amd-catalyst-windows-10/
    
    14.12 – ccc – Omega:
    http://www.guru3d.com/files-details/amd-catalyst-14-12-whql-omega-download.html
    
    13.3b3 – ccc – HD7790 release:
    http://www.guru3d.com/files-details/amd-catalyst-13-3-beta-2-download.html
    http://www.*****.com/?page=Files&go=giveme&dwn_id=1746
    
    12.11b11 – performance driver:
    http://www.guru3d.com/files-details/amd-catalyst-12-11-beta-download.html
    http://www.*****.com/?page=Files&go=giveme&dwn_id=1695
    
    12.2 – Stable GCN (12.x 8.96):
    http://www.guru3d.com/files-details/amd-catalyst-12-x-8-96-february-28.html
    
    11.12 – release GCN:
    http://www.guru3d.com/files-details/amd-catalyst-11-12-whql-8-921-5-dec-29-whql.html
    Links for 13.3b3/12.11b11 have '*****' which corresponds to 'ngo hq' without space in middle.
    - 2nd link is provided because local g3d link appeared to be broken.
    - only allowed modification is to add your card into supported set (to allow driver installation)

    UPDATE 2: Official set of Benchmarks (high availability).
    Get banchmarks, ready. Do not benchmark blindly without speadsheet to fill. (UPDATE3)
    Document has several sheets. Benchmarking participants are to fill only cells with yellow background. 'TestSetup' is one where you fill your info.
    Each benchmark has its own sheet.
    - And has 2 sets for different CPU clock (if you are into comparing impact of CPU as well). ThrottleStop can disable Turbo on intel's CPUs without need to reset system.
    - Cell A1 contains benchmark API to test. Cell B1 has link in case you need it.
    - Benchmarks have pre-filled 'Driver Package Version' and results as an example. Please replace those with your values.
    - Sheet 'Drivers' has set of links for driver download (just in case)

    IMPORTANT:

    - Once you decide on settings for given benchmark (resolution, details). Do not change it in older drivers even if it means bad performance.
    - We are testing to see differences drivers make on same hardware.

    UPDATE 3: OpenOffice spreadsheets
    OneDrive Shared Folder.
    - nVidia drivers are still not selected. Once I get those missing driver versions/links, I'll get you nVidia version (please, do not waste your time on random drivers).
     
    Last edited: Apr 16, 2016
  4. nz3777

    nz3777 Ancient Guru

    Messages:
    2,504
    Likes Received:
    215
    GPU:
    Gtx 980 Radeon 5500
    353.52 i am currently on 361.91 Wonder if I schould go back to old driver? What do you guys think? Or dont mess with it unless its broken type of deal?

    I have not noticed any performance loss out of my 780 thus-far,I mainly play Fallout 4 and Metro series games, with the occasional flight-sim Dcs world. I would be happy to help if you guys need it.

    i7 4790k @ 4.6 Ghz
    Gtx 780 sc @ 1163 Mhz core
     

  5. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    Why not FFXIV: Heavensward ?
     
  6. Undying

    Undying Ancient Guru

    Messages:
    25,332
    Likes Received:
    12,743
    GPU:
    XFX RX6800XT 16GB
    So what? We gonna bench some old synthetic tests and demos to deside which card performs better?

    Im interested in nz3777 gtx780 to compare my 280X. Let me know what kinda games do you own and if you are interested in some testing.
     
  7. -Tj-

    -Tj- Ancient Guru

    Messages:
    18,097
    Likes Received:
    2,603
    GPU:
    3080TI iChill Black
    Nv driver

    347.25? & 347.72 both icaffe
    353.50 or was it .52
    356.04
    362.00 (somewhat ok)
    364.72
    364.91 is ok too, need to run some tests again.
     
    Last edited: Apr 13, 2016
  8. moab600

    moab600 Ancient Guru

    Messages:
    6,658
    Likes Received:
    557
    GPU:
    PNY 4090 XLR8 24GB
    This thread is pointless ,kepler was left behind there wasn't any downgrade.

    AMD improved performance of GCN over time, while kepler stayed the same, the weakness of it exposed vs the strength of GCN.

    Maxwell will be in the same state., that's why AMD will always have the superior arch, GCN modification can be applied to lot of their products, tweaks and such applies to all GCN arch, that's why for them it easier to support this arch, rather than nvidia with stupid kepler then maxwell now pascal and volta.

    They should've tweaked Fermi, Fermi on 16nm would be such an awesome beast with the right tweaks.
     
  9. Fox2232

    Fox2232 Guest

    Messages:
    11,808
    Likes Received:
    3,371
    GPU:
    6900XT+AW@240Hz
    Sorry, my bad, fixed, I wrote it badly :)
    Thread may seem pointless, but it would be interesting to see how much. Or maybe tests will show something else what may be of interest even to you.
    Those are all very new drivers, we are looking for important drivers to cover entire Kepler Life span.
     
    Last edited: Apr 13, 2016
  10. Undying

    Undying Ancient Guru

    Messages:
    25,332
    Likes Received:
    12,743
    GPU:
    XFX RX6800XT 16GB
    Ok, this is my record for GCN1.0 and Tahiti chip. I broke 10k on graphics score and pushed my card to its limits. I would like to see how much are Kepler cards getting. Anyone with a 770/780 please leave your score and compare.

    [​IMG]
     
    Last edited: Apr 13, 2016

  11. Dragondale13

    Dragondale13 Ancient Guru

    Messages:
    1,527
    Likes Received:
    244
    GPU:
    GTX 1070 AMP! • H75
    331.93 beta for BF4 and 347.09 Titan drivers were the best drivers for Kepler in every game I played at the time.The performance boost from 353.06 has remained consistent up to 364.91 on my 670s but has introduced some minor streaming issues in games, mostly texture and shadow pop-in but not a big deal as the mainstream cards now have more vram, it's to be expected.
     
  12. -Tj-

    -Tj- Ancient Guru

    Messages:
    18,097
    Likes Received:
    2,603
    GPU:
    3080TI iChill Black
    Anything bellow 337.xx is not relevant anymore, nv made a boost since r337.. Kepler support is introduced since 320.xx or so, but anything bellow that is Fermi release territory (initial Fermi rls ~ 260.xx driver).




    I guess "Invalid score" is because of AMD tessellation hack? turn it off and try again :p


    @ 780GTX I got stock boost 1084MHZ gpu score ~10700.. OC'ed to ~ 1215mhz was ~11900, or ~12300 with image quality "hack" set to h.perf .
     
    Last edited: Apr 13, 2016
  13. -Tj-

    -Tj- Ancient Guru

    Messages:
    18,097
    Likes Received:
    2,603
    GPU:
    3080TI iChill Black
    double post..
     
  14. Undying

    Undying Ancient Guru

    Messages:
    25,332
    Likes Received:
    12,743
    GPU:
    XFX RX6800XT 16GB
    Slider is set to 16X. It makes a difference in performance with no visual quality lost. No one said nvidia to not include the slider, curse them not me. :p

    Btw, 280X = GTX780 stock as you can see. 770 cant even get close to these scores.
     
    Last edited: Apr 13, 2016
  15. Fox2232

    Fox2232 Guest

    Messages:
    11,808
    Likes Received:
    3,371
    GPU:
    6900XT+AW@240Hz
    Well, available information places 1st Kepler desktop card at 2012-03-22, which fits HD7970 release date (bit earlier).
    I went here to get matching dates/drivers.
    And according to HH 301.10 is GTX680 release driver. So that's one solid checkpoint.
    353.51 is supposedly in performance improvement category. (-tj-, Dragon) So, it is 2nd to use.
    1st Fermi (GTX 480) released 2010/03/26, oldest g3d driver mirror: 2010/05/18, so that can be Fermi starting point.

    Now, I'll update this post with dates. And where we should look for other drivers.
    <Newest Driver Here>
    353.51 - 2015/07/18
    xXx.Xx - (2013/06 ~ 2014/12)
    xXx.Xx - (2012/10 ~ 2013/01)
    301.10 - 2012/03/23
    197.75 - 2010/05/18 - Fermi (Does not include Kepler)

    Date ranges are limited in way that we'll have driver before and after clusters of games/benchmarks. So we'll see at least some indication of per game improvement vs. 'over time cumulative' improvement.
     
    Last edited: Apr 13, 2016

  16. WhiteLightning

    WhiteLightning Don Illuminati Staff Member

    Messages:
    30,766
    Likes Received:
    3,934
    GPU:
    Inno3d RTX4070
    lol @ using hack to get better score :D
     
  17. Ieldra

    Ieldra Banned

    Messages:
    3,490
    Likes Received:
    0
    GPU:
    GTX 980Ti G1 1500/8000
    Yeah, but you're making up for low geometry performance by doing that, there is a difference in IQ but it's subtle, this complicates testing.

    Hypothetical situation

    1. 280x with 16x tess doesn't hit its geometry bottleneck
    2. 780 with default tess hits geometry bottleneck

    This is now an unfair test, how do we standardize this

    @Tj you mean the IQ settings in nvcp?
     
    Last edited: Apr 13, 2016
  18. Fox2232

    Fox2232 Guest

    Messages:
    11,808
    Likes Received:
    3,371
    GPU:
    6900XT+AW@240Hz
    Guys, this is not about measuring your eP, or AMD vs nVidia. Each and every participant will get set of tests, and because we can't guarantee same testing systems even for AMD vs AMD. All results will be normalized. And basically each participant is competing with himself. (In other words no competition at all. It is just driver vs game comparison with relatively decent granularity over time.)

    It will make nice graphs and tables, I already got few ideas how to visualize in way it will make sense to everyone.
     
  19. vase

    vase Guest

    Messages:
    1,652
    Likes Received:
    2
    GPU:
    -
    gimme gimme!
     
  20. -Tj-

    -Tj- Ancient Guru

    Messages:
    18,097
    Likes Received:
    2,603
    GPU:
    3080TI iChill Black
    Ieldra

    Yes. And slight lod+ adjustment in nvinspector, cowie told me about this trick once for hwbot benchmarking.

    Although I havent tested this yet on 980ti :nerd:


    Fox
    Those are initial drivers, well i listed more "mature" drivers, btw fermi had a 191, 195 jump straight to 256 in next rls.. also I was only focused on full/big fermi & kepler gpus.. anyway no point testing initial drivers because they're also buggy by newer games. I dare to say it will be the same by AMD initial rls drivers.

    I dont curse anyone, its just that score is fake. So it doesnt come close to anything unless you re-test with tessellation override off :nerd:
     
    Last edited: Apr 14, 2016

Share This Page