Synchroniszation issues with SLI and CrossFire

[tombman]

AFR will never have 100% identical frame distribution.

Hi, i'm the original author of the June 2007 forum post at 3dcenter.de.

I contacted tomshardware, fudzilla and anandtech with this AFR-problem but they didn't react. Obviously they don't want to piss Nvidia off because telling the world+dog that SLI is only really usable at framerates at 45fps and beyond is not what Nvidia wants to hear

And please don't play down the problem, as it IS a big one.
It questions the sense of sli at a whole. When you get 20fps with a single card, and you consider buying a second card to get over 30fps by using SLI, and the subjective impression of fluidness of the pictures is NOT improved, then SLi gets senseless in general!!!

Usually people get SLI because they want to run quality settings impossible for a single card. That means: the framerate range at around 30fps to 40fps. And that is exactly the range where the AFR problem is most severe!!!

And, what you don't tell the people, there IS a solution to this problem- a CONTROLLED DELAY OF FRAME OUTPUT.
The driver could MEASURE how long it took to render a frame and then estimate an ADDED WAITING TIME for the next frame to be ouput.

Enthusiasts have very fast cpus, and the less time the cpu needs to prepare a frame and the more time the gpus take to render it the worse the AFR problem gets.
Within 20ms the cpu prepares 2 frames and sends them to the gpus, so there is only a 10ms delay between both gpus rendering on their frames. So when each gpu needs 50ms for each frame you will see something like this:

10ms-50ms-10ms-50ms -> fast, slow, fast, slow

The human eye will allmost competely ignore the fast 10ms ones (=100fps), but will only see the slow 50ms ones. And 50ms means 20fps -> THE SAME "OPTICAL" FRAMERATE AS WITH A SINGLE CARD. -> WASTED MONEY.

You tell the people "there is no solution...", but in that sentence there is something missing, that is: "...WITHOUT SACRIFICING A SMALL AMOUNT OF AVERAGE FPS"

And losing fps, albeit very little, means "ATI will look good" which means "Nvidia will look bad". Obviously ATI is thinking the same, as they have the SAME problem with Crossfire.

So none of the two implements an intelligent, adaptable delay because it means a small amout of average fps loss

An intelligent delay would NOT ALLOW a frame to be ouput only 10ms after the previous one, knowing that rendering the previous frame has taken 50ms! It would delay it for another 20ms and THEN put it out. What would happen?
THIS: 30ms-30ms-30ms-30ms -> perfect distribution = perfect 30fps -> problem solved

Why intelligent? Because, when the framerate gets faster, which means the computational load gets lighter, the delay would get much smaller, it would ADAPT.
Many adaptions mean bigger loss of average fps, BUT, and this is important, content doesn't change that fast in a game.
In the world of microseconds changes in content happen SLOWLY.

I don't think the average fps would suffer more than 5%. But, VISUALLY it would be a difference like night and day.

I have tested Half Life 2, which has an own fps limiter built into the engine, that can be activated via the console.
I used settings where i got slightly over 30fps WITH SLI TURNED ON. Then i measured the frametimes and the distribution was horrible.
And now comes the fun part: i activated the INTERNAL HL2 FPS Limiter at exactly 30fps and BAM-> perfect distribution.

So, what was the verdict? It IS possible to get perfect distribution of frametime WITH SLI, when you use some kind of limiter/delay.
Obviously the best limiter, the one that can react the quickest, can only be implemented inside the DRIVER, which means: it is Nvidia's task to code one.

I'm sure that it is not only me that would LOVE to sacrifice 5%, let it be 10%, of average fps when, in return, a perfect distribution of frametimes can be achieved. (i'm a SLI user since day 1)

p.s. sorry for my bad english.

 

[apoppin]

There is no CF filtering,sigh. SuperAA is SuperAA. Or is the filtering thing like the SMIC one or the Tesla 2.0 one?In which case i'm sorry for having interfered.

You are beginning to annoy me - very slightly - also


Well, AA *IS* strictly speaking a filter, different from texture filtering, but still a filter. Why else would ATi insist that CFAA stands for Custom FILTER Anti Aliasing?

AND the gentleman poster above me is quite correct; and MS is fixable imo.

 

[tombman]

Since i may not edit my post: there is a small error in my post "it should say 60ms for each gpu, not 50ms"...

 

[ChrisRay]

The sacrafice is higher than 10% FPS. You dont dynamically balance and sync the frames. Your framerate will eventually drop below that of a single card. MS delays that spike are typically a readjustment of load balancing. If a game is horrible at load balancing. The problem becomes larger.

I know everyone here thinks they know more about the issue than the guys at Nvidia/AMD. But if the solution was as simple as you make it out to be. It would have been done a long time ago.

But if you wish to vent at me. Say Im downplaying something I consider relatively a minor issue. Go ahead. I've gotten quite used to it.

 

[SirPauly]

There's no doubt that nVidia and ATI know much more about 3d but they don't go out of there way to discuss limitations or cons too much ChrisRay.

My findings are very similar to what Tombman is offering and have shared this view years ago -- at around 40-45 frames is where the fine-line may be. If gamers didn't vent or share their views how would anyone even know?

Most web-sites doing reviews? Hardly! Their gaming experience for the most part is benching and offering PR material. It would be so refreshing to read investigative articles on the gaming experience instead of cookie cutters.

Smoothness is still subjective to each individual gamer though..........and why views differ so. And AFR is still so very welcomed if it's placed in an objective context.

 

[compres]

The sacrafice is higher than 10% FPS. You dont dynamically balance and sync the frames. Your framerate will eventually drop below that of a single card. MS delays that spike are typically a readjustment of load balancing. If a game is horrible at load balancing. The problem becomes larger.

I know everyone here thinks they know more about the issue than the guys at Nvidia/AMD. But if the solution was as simple as you make it out to be. It would have been done a long time ago.

But if you wish to vent at me. Say Im downplaying something I consider relatively a minor issue. Go ahead. I've gotten quite used to it.

The thing is most people here find this problem annoying or are at least curious about it and you keep saying it is a "no issue", while maintaining a very biased attitude towards nVidia SLI.

I don't think anyone here is trying to attack you personally.

 

[ChrisRay]

I have said and acknowledged the issue countless times. I have even provided feedback to some users at slizone on this issue ((After it was brought up here)) on what to look for. If you want me to agree that its some "Earth Shattering Western society overlook". Then I wont. I just dont think its all that big of a deal.

I sum it up as a mountain out of an ant hill. And thats what it is to me. Everyone seems to think there is some "easy" solution to the problem. But their's not. You try to force a constant frame sync from SLI. Then you'll eventually lose all benefits of SLI. It's the entire point and purpose of dynamic load balancing. There are going to be people who hate AFR. And they are entitled to their opinion. I'm not even gonna try to convince them otherwise. But I am not gonna raise some big stink over something thats been here for 3 + years. Especially when it doesn't even bother me personally.

 

[SirPauly]

That montain out of an ant hill is one of the reasons for craving a more powerful GPU than an AFR solution offering similar performance attributes.

 

[ChrisRay]

And to that extent. I agree. A single GPU offering the same performance as an AFR solution is better. Not because of profiles,. Because the performance is more consistent. I believe I've said that as well.

I just I dont see this as a monsterous problem and that all SLI/Crossfire users have been screwed. AFR can offer some legitimate performance gains. And lets me do things that no single GPU can do. Some people don't like side effects that can occasionally come with it. Thats perfectly fine. I certainly don't mind people taking issue with it. SLI has never been for everyone.

Chris

 

[annihilator]

Actually the recorded benchmark I use was just the one I use for all my UT3 Benchmarking. So yes the frame counter goes much further. As does rendered frames. I thought it was pretty obvious I wasnt including the whole thing because the post would have been a mile long of frame spam. There's nothing paticularly malicious about that. Other than trying to save some forum space for the few people who really care to read through all that. If I included all the frames. I would have had to need several posts worth of character space. All I was doing was demonstrating the typical frame distribution in UT3 and the overall results I got without spamming the forums in mile long page of data.

Like I said before. If anyone thins they can do a better job of obtaining a 100% accurate benchmark with the exact amount of frames rendered in a given sec. Then by all means. Be my guest. I've certainly got no problem with that. I am not dedicating that amount of time to something I consider pretty much a non issue. I have much bigger things on my table right now to deal with hardware wise.

Apoppin. SLIAA will behave like a single GPU in frame syncs.

It's amusing that in a whole paragraph you said absolutely nothing and just talked about why you omitted the frames from 100 and on. As I wrote above, it was pretty obvious you didn't include pages of frame times and I didn't criticize that. However, I criticize that even in your examples, in 2-way SLI the frame time synchronization sucks awesomely; and 3-way and 4-way don't even increase FPS in the part where it matters (where the FPS's are lowest).

 

[ChrisRay]

I'm glad you find it so funny. Considering that specific aspect was the only part of your post I did bother responding too. If your so convinced my data is flawed. Then be my guest and try to do your own comparative analysis with UT3. I think you'll find that synchronization for the most remains pretty constant regardless of what frame your actually "viewing" on the map I am testing. What I find so obnoxious is I'm the one being called biased and flamed yet I'm the only one here who has provided any real data on the subject and have done the damned best I could to describe the conditions of my benchmarks.

Here's another one I took of 3 way with BF2142 in the berlin map. Where the average frame was "88 FPS" ((BF2142 FPS was unlocked to 150 but it didnt really make a difference)) it also was the entirety of the benchmark. There 1 or 2 spikes ((40 MSish I noticed)) but to track the data over a long course of a benchmark would just be impossible time inducing task that I'm not really interested in doing. 63-64 is a noticable frame. Non of these frames has "perfect" distribution however. But the jumps are very minor. I chose this benchmark because its one of the "less ideal" situations I found. Where texture memory leaked out of the framebuffer and caused PCIE memory access. Which also can cause "stutters". It also uses different profile bits than than UT3 which makes it behave slightly different. Mode used was 16xS so it was adequately GPU bound.

Frame, Time (ms)
    1,     0.000
    2,    13.207
    3,    21.562
    4,    31.242
    5,    50.923
    6,    59.008
    7,    73.712
    8,    81.891
    9,    91.375
   10,   106.143
   11,   113.815
   12,   123.942
   13,   140.403
   14,   152.083
   15,   159.910
   16,   175.153
   17,   183.174
   18,   193.184
   19,   207.853
   20,   216.148
   21,   225.593
   22,   241.155
   23,   248.925
   24,   259.016
   25,   274.088
   26,   282.157
   27,   291.686
   28,   307.049
   29,   314.888
   30,   325.016
   31,   354.253
   32,   379.014
   33,   393.961
   34,   407.546
   35,   415.554
   36,   425.700
   37,   440.289
   38,   448.483
   39,   458.283
   40,   472.858
   41,   481.215
   42,   490.906
   43,   505.756
   44,   513.980
   45,   523.794
   46,   538.623
   47,   546.279
   48,   556.330
   49,   570.243
   50,   586.966
   51,   595.451
   52,   611.807
   53,   619.605
   54,   630.037
   55,   644.244
   56,   652.672
   57,   662.478
   58,   677.316
   59,   685.722
   60,   695.423
   61,   709.754
   62,   717.582
   63,   727.620
   64,   757.597
   65,   765.515
   66,   780.142
   67,   804.811
   68,   816.837
   69,   824.450
   70,   840.213
   71,   847.866
   72,   858.127
   73,   872.223
   74,   880.085
   75,   889.997
   76,   904.149
   77,   919.801
   78,   927.953
   79,   942.031
   80,   950.080
   81,   960.177
   82,   974.236
   83,   997.941
   84,  1010.456
   85,  1018.366
   86,  1028.682
   87,  1042.659
   88,  1050.806
   89,  1073.011
   90,  1086.436
   91,  1093.879
   92,  1108.925
   93,  1116.629
   94,  1126.661
   95,  1141.451
   96,  1149.581
   97,  1161.320
   98,  1174.318
   99,  1182.413
  100,  1192.177
  101,  1207.044
  102,  1215.179
  103,  1225.853
  104,  1239.079
  105,  1246.901
  106,  1272.847
  107,  1286.352
  108,  1294.518
  109,  1331.403
  110,  1343.982
  111,  1351.860
  112,  1362.097
  113,  1376.629
  114,  1384.734
  115,  1408.732
  116,  1421.268
  117,  1428.845
  118,  1453.452
  119,  1461.770
  120,  1476.494
  121,  1484.376
  122,  1494.390
  123,  1509.036
  124,  1517.051
  125,  1526.939
  126,  1543.296
  127,  1551.161
  128,  1561.176
  129,  1575.863
  130,  1584.028
  131,  1593.645
  132,  1609.597
  133,  1617.458
  134,  1627.682
  135,  1642.400
  136,  1650.525
  137,  1660.080
  138,  1674.860
  139,  1682.860
  140,  1692.636
  141,  1706.930
  142,  1714.925
  143,  1724.362
  144,  1740.384
  145,  1748.636
  146,  1758.265
  147,  1773.850
  148,  1781.928
  149,  1791.514
  150,  1806.579
  151,  1814.171
  152,  1824.335
  153,  1838.828
  154,  1846.865
  155,  1856.360
  156,  1871.062
  157,  1878.664
  158,  1888.905
  159,  1898.631
  160,  1913.990
  161,  1921.396
  162,  1931.860
  163,  1945.840
  164,  1953.657
  165,  1963.214
  166,  1978.358
  167,  1985.790
  168,  1995.963
  169,  2012.160
  170,  2020.059
  171,  2029.807
  172,  2044.175
  173,  2051.692
  174,  2061.788
  175,  2076.551
  176,  2084.499
  177,  2093.610
  178,  2109.329
  179,  2116.320
  180,  2126.487
  181,  2140.503
  182,  2148.791
  183,  2158.282
  184,  2174.268
  185,  2182.388
  186,  2192.042
  187,  2206.076
  188,  2213.747
  189,  2223.510
  190,  2241.076
  191,  2248.379
  192,  2258.539
  193,  2272.738
  194,  2280.522
  195,  2290.050
  196,  2305.360
  197,  2312.649
  198,  2322.895
  199,  2338.462
  200,  2346.815
  201,  2355.969
  202,  2370.030
  203,  2377.373
  204,  2387.454
  205,  2397.335
  206,  2412.261
  207,  2419.505
  208,  2429.683
  209,  2443.749
  210,  2451.540
  211,  2460.698
  212,  2476.556
  213,  2483.582
  214,  2493.639
  215,  2508.752
  216,  2516.148
  217,  2525.759
  218,  2540.174
  219,  2547.168
  220,  2557.312
  221,  2571.705
  222,  2579.294
  223,  2588.937
  224,  2603.383
  225,  2610.627
  226,  2620.811
  227,  2630.718
  228,  2645.455
  229,  2652.917
  230,  2662.820
  231,  2677.688
  232,  2685.272
  233,  2694.956
  234,  2711.645
  235,  2718.998
  236,  2728.941
  237,  2743.312
  238,  2751.105
  239,  2760.689
  240,  2775.827
  241,  2783.186
  242,  2793.382
  243,  2808.620
  244,  2816.334
  245,  2825.909
  246,  2840.410
  247,  2847.382
  248,  2857.583
  249,  2871.730
  250,  2879.043
  251,  2888.805
  252,  2903.611
  253,  2910.676
  254,  2920.765
  255,  2930.877
  256,  2946.874
  257,  2954.817
  258,  2964.631
  259,  2979.195
  260,  2986.829
  261,  2996.367
  262,  3012.040
  263,  3019.485
  264,  3029.810
  265,  3044.178
  266,  3051.997
  267,  3061.679
  268,  3076.130
  269,  3083.601
  270,  3093.829
  271,  3109.365
  272,  3117.306
  273,  3126.938
  274,  3142.113
  275,  3149.707
  276,  3159.745
  277,  3175.827
  278,  3183.687
  279,  3193.288
  280,  3209.057
  281,  3216.548
  282,  3226.825
  283,  3242.569
  284,  3252.349
  285,  3261.872
  286,  3278.012
  287,  3289.041
  288,  3297.522
  289,  3312.911
  290,  3321.909
  291,  3330.690
  292,  3345.410
  293,  3363.201
  294,  3376.079
  295,  3383.486
  296,  3393.762
  297,  3410.300
  298,  3418.169
  299,  3427.766
  300,  3442.542
  301,  3466.937
  302,  3479.795
  303,  3487.385
  304,  3497.605
  305,  3512.465
  306,  3520.457
  307,  3529.752
  308,  3545.535
  309,  3553.072
  310,  3563.205
  311,  3577.390
  312,  3585.160
  313,  3594.590
  314,  3610.258
  315,  3617.327
  316,  3627.550
  317,  3643.500
  318,  3651.827
  319,  3660.691
  320,  3675.650
  321,  3683.101
  322,  3692.866
  323,  3708.103
  324,  3716.288
  325,  3726.043
  326,  3740.991
  327,  3748.156
  328,  3758.746
  329,  3772.103
  330,  3779.510
  331,  3789.080
  332,  3804.181
  333,  3811.231
  334,  3821.453
  335,  3831.280
  336,  3846.528
  337,  3853.476
  338,  3863.683
  339,  3880.209
  340,  3888.151
  341,  3897.979
  342,  3913.044
  343,  3921.093
  344,  3930.519
  345,  3944.612
  346,  3952.142
  347,  3961.821
  348,  3977.093
  349,  3984.327
  350,  3994.490
  351,  4009.102
  352,  4016.756
  353,  4026.349
  354,  4041.306
  355,  4048.607
  356,  4058.759
  357,  4073.851
  358,  4081.679
  359,  4091.015
  360,  4108.536
  361,  4116.019
  362,  4126.031
  363,  4140.749
  364,  4149.146
  365,  4158.166
  366,  4175.292
  367,  4182.598
  368,  4192.903
  369,  4208.429
  370,  4216.319
  371,  4225.799
  372,  4241.170
  373,  4248.523
  374,  4258.767
  375,  4272.898
  376,  4280.576
  377,  4290.150
  378,  4305.437
  379,  4312.663
  380,  4322.867
  381,  4339.258
  382,  4346.992
  383,  4356.573
  384,  4372.633
  385,  4380.215
  386,  4390.157
  387,  4405.228
  388,  4413.173
  389,  4422.460
  390,  4438.278
  391,  4445.715
  392,  4455.956
  393,  4470.217
  394,  4477.846
  395,  4487.280
  396,  4497.691
  397,  4512.262
  398,  4519.890
  399,  4529.428
  400,  4544.445
  401,  4551.578
  402,  4561.791
  403,  4578.385
  404,  4585.846
  405,  4595.239
  406,  4610.563
  407,  4617.460
  408,  4627.757
  409,  4641.657
  410,  4649.187
  411,  4658.881
  412,  4676.141
  413,  4683.628
  414,  4693.259
  415,  4708.976
  416,  4716.902
  417,  4726.242
  418,  4742.269
  419,  4749.304
  420,  4759.412
  421,  4774.609
  422,  4781.911
  423,  4791.412
  424,  4808.678
  425,  4815.440
  426,  4825.753
  427,  4842.062
  428,  4850.041
  429,  4859.059
  430,  4875.808
  431,  4883.471
  432,  4893.354
  433,  4911.286
  434,  4919.299
  435,  4928.473
  436,  4944.847
  437,  4952.011
  438,  4962.102
  439,  4977.338
  440,  4984.965
  441,  4994.696
  442,  5010.638
  443,  5017.721
  444,  5027.969
  445,  5045.139
  446,  5052.712
  447,  5062.159
  448,  5077.917
  449,  5085.070
  450,  5095.113
  451,  5110.793
  452,  5118.002
  453,  5127.593
  454,  5143.152
  455,  5149.995
  456,  5160.164
  457,  5175.789
  458,  5183.751
  459,  5192.824
  460,  5209.599
  461,  5216.558
  462,  5226.979
  463,  5242.157
  464,  5249.671
  465,  5259.120
  466,  5278.851
  467,  5286.041
  468,  5296.140
  469,  5311.834
  470,  5319.404
  471,  5328.879
  472,  5344.912
  473,  5352.024
  474,  5362.163
  475,  5378.159
  476,  5385.691
  477,  5395.082
  478,  5411.957
  479,  5419.028
  480,  5429.215
  481,  5444.711
  482,  5452.260
  483,  5461.696
  484,  5478.097
  485,  5485.163
  486,  5495.270
  487,  5512.403
  488,  5519.943
  489,  5530.086
  490,  5544.996
  491,  5552.124
  492,  5565.478
  493,  5576.915
  494,  5584.313
  495,  5594.585
  496,  5610.152
  497,  5616.904
  498,  5627.378
  499,  5650.827
  500,  5658.428
  501,  5672.298
  502,  5680.073
  503,  5689.186
  504,  5705.430
  505,  5712.950
  506,  5722.827
  507,  5739.254
  508,  5747.204
  509,  5756.684
  510,  5773.956
  511,  5781.093
  512,  5792.106
  513,  5812.632
  514,  5820.936
  515,  5829.741
  516,  5844.707
  517,  5851.395
  518,  5861.627
  519,  5876.346
  520,  5883.765
  521,  5906.107
  522,  5918.162
  523,  5925.010
  524,  5940.024
  525,  5946.993
  526,  5957.117
  527,  5973.293
  528,  5980.562
  529,  5990.308
  530,  6004.829
  531,  6011.531
  532,  6021.944
  533,  6037.341
  534,  6045.056
  535,  6054.191
  536,  6065.108
  537,  6079.518
  538,  6086.635
  539,  6096.248
  540,  6111.620
  541,  6118.338
  542,  6128.667
  543,  6143.327
  544,  6150.513
  545,  6160.108
  546,  6175.092
  547,  6181.926
  548,  6192.247
  549,  6208.920
  550,  6216.274
  551,  6225.701
  552,  6241.376
  553,  6248.352
  554,  6258.412
  555,  6272.886
  556,  6280.125
  557,  6289.665
  558,  6305.224
  559,  6332.148
  560,  6344.899
  561,  6351.701
  562,  6362.088
  563,  6375.997
  564,  6383.366
  565,  6393.043
  566,  6410.000
  567,  6417.661
  568,  6427.605
  569,  6444.719
  570,  6452.663
  571,  6461.677
  572,  6478.108
  573,  6485.154
  574,  6495.203
  575,  6510.525
  576,  6518.009
  577,  6527.677
  578,  6543.095
  579,  6550.353
  580,  6560.365
  581,  6575.348
  582,  6582.451
  583,  6591.949
  584,  6613.113
  585,  6620.401
  586,  6643.414
  587,  6654.740
  588,  6661.685
  589,  6677.620
  590,  6684.601
  591,  6694.365
  592,  6710.877
  593,  6718.236
  594,  6728.065
  595,  6742.696
  596,  6749.924
  597,  6759.855
  598,  6776.486
  599,  6783.490
  600,  6793.469
  601,  6808.718
  602,  6816.135
  603,  6825.578
  604,  6841.109
  605,  6848.072
  606,  6858.128
  607,  6872.831
  608,  6880.189
  609,  6889.880
  610,  6908.573
  611,  6916.215
  612,  6926.396
  613,  6941.636
  614,  6949.064
  615,  6958.942
  616,  6981.842
  617,  6988.872
  618,  7004.784
  619,  7011.533
  620,  7021.669
  621,  7036.108
  622,  7043.081
  623,  7052.684
  624,  7062.983
  625,  7077.657
  626,  7084.721
  627,  7094.353
  628,  7110.317
  629,  7116.920
  630,  7127.465
  631,  7157.080
  632,  7164.687
  633,  7178.823
  634,  7186.033
  635,  7195.601
  636,  7210.730
  637,  7217.449
  638,  7227.645
  639,  7242.311
  640,  7249.454
  641,  7258.959
  642,  7274.693
  643,  7281.599
  644,  7291.698
  645,  7306.205
  646,  7313.477
  647,  7323.100
  648,  7338.700
  649,  7345.210
  650,  7355.409
  651,  7371.162
  652,  7378.057
  653,  7387.668
  654,  7397.860
  655,  7412.600
  656,  7419.538
  657,  7429.169
  658,  7443.695
  659,  7457.982
  660,  7470.473
  661,  7477.280
  662,  7487.345
  663,  7497.168
  664,  7512.552
  665,  7519.341
  666,  7529.473
  667,  7556.777
  668,  7563.930
  669,  7577.891
  670,  7585.026
  671,  7594.726
  672,  7611.990
  673,  7619.030
  674,  7629.061
  675,  7643.943
  676,  7651.183
  677,  7660.803
  678,  7675.810
  679,  7682.630
  680,  7692.757
  681,  7708.484
  682,  7715.789
  683,  7725.081
  684,  7740.134
  685,  7747.538
  686,  7757.180
  687,  7772.834
  688,  7780.078
  689,  7789.498
  690,  7805.017
  691,  7812.004
  692,  7821.807
  693,  7838.156
  694,  7845.095
  695,  7854.938
  696,  7870.718
  697,  7877.235
  698,  7897.962
  699,  7910.169
  700,  7917.123
  701,  7926.682
  702,  7941.500
  703,  7947.948
  704,  7958.227
  705,  7972.692
  706,  7979.490
  707,  7989.109
  708,  8003.976
  709,  8010.521
  710,  8020.730
  711,  8030.600
  712,  8046.173
  713,  8054.504
  714,  8062.907
  715,  8080.401
  716,  8088.582
  717,  8097.002
  718,  8112.731
  719,  8119.290
  720,  8130.815
  721,  8146.754
  722,  8153.758
  723,  8166.283
  724,  8179.026
  725,  8195.399
  726,  8208.073
  727,  8214.668
  728,  8236.551
  729,  8248.041
  730,  8255.079
  731,  8275.746
  732,  8282.754
  733,  8292.260
  734,  8309.381
  735,  8315.887
  736,  8326.028
  737,  8340.773
  738,  8347.785
  739,  8357.231
  740,  8372.588
  741,  8379.166
  742,  8389.308
  743,  8403.386
  744,  8410.614
  745,  8419.928
  746,  8430.561
  747,  8444.903
  748,  8451.926
  749,  8461.470
  750,  8477.180
  751,  8483.819
  752,  8493.965
  753,  8508.719
  754,  8515.781
  755,  8525.537
  756,  8541.906
  757,  8548.423
  758,  8558.641
  759,  8573.462
  760,  8580.591
  761,  8590.175
  762,  8605.549
  763,  8612.390
  764,  8622.509
  765,  8648.408
  766,  8655.806
  767,  8670.399
  768,  8677.733
  769,  8687.237
  770,  8697.048
  771,  8711.447
  772,  8718.193
  773,  8727.852
  774,  8743.329
  775,  8749.889
  776,  8760.343
  777,  8777.503
  778,  8784.497
  779,  8794.070
  780,  8809.550
  781,  8816.488
  782,  8826.848

 

[annihilator]

Where did I ever say your data is "flawed"? I don't doubt its accuracy. The point I didn't get is how you could defend that SLI is a nice thing to have, looking at those figures. The 2-Way frame time you posted there isn't very nicely synchronised, at all.

If providing data makes you right, I can provide tons of games to you in which there isn't any "syhchronisation" at all; as well.

Here's one I just did.

3DMark06 (1680x1050 8xAA 16xAF)

Nice, right? Frame time differences: 19-45-17-47-18-50 etc. So the benchmark shows that I'm seeing 30/0.962 = 32 FPS, whereas one out of every two frame I see is rendered in a 50 ms delay, which makes the practical FPS only 20. And it's a stuttering 20 FPS.

Disabling SLI gives the following results:

All the frames have perfectly the same delay. Looking at FPS: 30/1.46 = 20.5FPS. And it looks way smoother than SLI, due to the balanced delay.

So yeah, let every review site blindly publish SLI bench numbers as if they were anywhere near comparable with non-AFR rendering. Let them say that "by adding a second card our FPS bumped from 20 to 32, a %65 increase in performance!!" whereas the performance actually LOWERED and the viewing experience visibly WORSENED, let alone increasing %65 or even staying the same.

 

[ChrisRay]

The point I didn't get is how you could defend that SLI is a nice thing to have, looking at those figures. The 2-Way frame time you posted there isn't very nicely synchronised, at all.

Because I like it? Because it lets me raise AA/AF and get playable framerates in games I play? You say that AFR rendering provides no benefit. I completely disagree. AFR allows for performance and IQ to be obtained that a single GPU may not be able to provide. I have never once expressed that the synchronization would be perfect. I said the MS delays and intermittent changes are largely un noticable. And they are in most of the games I play. When I see large MS delays, They are far more noticable than small MS delays with minor variations in synchronization. The 10-15 MS fluctuations are pretty much un noticable. Its the large ones that change from 50-100 MS that completely undesired. 15-25-33-45-120-129-139-152 but they just dont happen that often and are often explained by other problems.

Heres another example where I deliberately went out of my way to cause the frames to desync in Neverwinter Nights 2. Widly changing the camera range which caused the load balancer to "flip" out occasionally. And even in this case. The synchronization didnt go too far out of whack and quickly corrected itself over the course of 400 frames with an AVG of 61 FPS. The most distracting thing you can see is when the load balancer fails a prediction and has a large MS delay descynchronization.

Frame, Time (ms)
    1,     0.000
    2,    14.990
    3,    29.806
    4,    44.748
    5,    59.423
    6,    74.186
    7,    88.837
    8,   103.644
    9,   118.534
   10,   133.522
   11,   175.104
   12,   192.136
   13,   208.883
   14,   237.842
   15,   254.878
   16,   271.980
   17,   289.057
   18,   305.962
   19,   352.952
   20,   370.813
   21,   387.477
   22,   406.171
   23,   424.725
   24,   442.010
   25,   459.214
   26,   476.621
   27,   494.135
   28,   511.218
   29,   528.321
   30,   545.077
   31,   562.356
   32,   578.972
   33,   596.094
   34,   613.126
   35,   650.331
   36,   668.987
   37,   688.498
   38,   704.934
   39,   721.873
   40,   738.950
   41,   783.873
   42,   802.653
   43,   820.240
   44,   837.269
   45,   854.116
   46,   874.059
   47,   891.997
   48,   910.310
   49,   932.534
   50,   952.371
   51,   969.367
   52,   984.801
   53,  1000.214
   54,  1015.403
   55,  1032.753
   56,  1050.192
   57,  1066.896
   58,  1102.084
   59,  1118.625
   60,  1135.754
   61,  1152.568
   62,  1177.397
   63,  1194.684
   64,  1211.651
   65,  1228.211
   66,  1244.610
   67,  1259.512
   68,  1286.786
   69,  1303.317
   70,  1319.628
   71,  1336.215
   72,  1352.347
   73,  1368.485
   74,  1392.132
   75,  1409.797
   76,  1426.244
   77,  1442.839
   78,  1458.955
   79,  1475.022
   80,  1492.121
   81,  1508.935
   82,  1525.267
   83,  1541.452
   84,  1557.763
   85,  1574.211
   86,  1590.454
   87,  1606.813
   88,  1622.992
   89,  1639.261
   90,  1655.421
   91,  1673.035
   92,  1689.335
   93,  1705.420
   94,  1721.582
   95,  1737.854
   96,  1755.796
   97,  1772.130
   98,  1788.485
   99,  1804.666
  100,  1820.911
  101,  1837.863
  102,  1854.297
  103,  1870.565
  104,  1887.800
  105,  1904.736
  106,  1921.424
  107,  1938.169
  108,  1967.110
  109,  1984.513
  110,  2002.017
  111,  2019.258
  112,  2036.199
  113,  2052.957
  114,  2069.117
  115,  2086.672
  116,  2103.387
  117,  2119.783
  118,  2136.757
  119,  2151.328
  120,  2170.998
  121,  2189.529
  122,  2205.712
  123,  2221.952
  124,  2238.123
  125,  2255.243
  126,  2270.025
  127,  2284.347
  128,  2298.684
  129,  2315.023
  130,  2332.081
  131,  2348.965
  132,  2365.499
  133,  2381.923
  134,  2398.492
  135,  2415.080
  136,  2431.616
  137,  2456.921
  138,  2473.457
  139,  2490.034
  140,  2506.309
  141,  2523.029
  142,  2539.444
  143,  2556.031
  144,  2572.718
  145,  2621.604
  146,  2638.089
  147,  2654.438
  148,  2673.617
  149,  2689.964
  150,  2706.664
  151,  2723.131
  152,  2739.470
  153,  2756.289
  154,  2772.293
  155,  2790.379
  156,  2811.977
  157,  2839.932
  158,  2856.643
  159,  2873.067
  160,  2889.192
  161,  2905.912
  162,  2922.184
  163,  2938.346
  164,  2955.058
  165,  2971.288
  166,  2987.474
  167,  3003.563
  168,  3020.032
  169,  3036.449
  170,  3052.585
  171,  3068.793
  172,  3085.286
  173,  3101.547
  174,  3149.068
  175,  3165.344
  176,  3181.546
  177,  3197.680
  178,  3216.365
  179,  3232.574
  180,  3248.714
  181,  3264.992
  182,  3281.110
  183,  3297.064
  184,  3313.560
  185,  3329.752
  186,  3346.459
  187,  3362.897
  188,  3379.032
  189,  3395.681
  190,  3411.820
  191,  3428.077
  192,  3444.609
  193,  3461.203
  194,  3477.588
  195,  3493.696
  196,  3519.839
  197,  3536.022
  198,  3552.464
  199,  3572.152
  200,  3588.292
  201,  3604.678
  202,  3620.819
  203,  3637.175
  204,  3653.374
  205,  3672.038
  206,  3688.288
  207,  3704.337
  208,  3720.715
  209,  3736.772
  210,  3752.815
  211,  3768.957
  212,  3790.189
  213,  3807.929
  214,  3824.189
  215,  3840.432
  216,  3856.430
  217,  3872.384
  218,  3888.352
  219,  3904.753
  220,  3920.798
  221,  3937.490
  222,  3954.058
  223,  3970.397
  224,  3986.673
  225,  4002.707
  226,  4019.283
  227,  4035.580
  228,  4052.002
  229,  4068.015
  230,  4084.278
  231,  4100.350
  232,  4116.224
  233,  4132.121
  234,  4147.847
  235,  4164.060
  236,  4183.946
  237,  4199.828
  238,  4220.151
  239,  4261.038
  240,  4277.101
  241,  4292.884
  242,  4308.639
  243,  4324.337
  244,  4340.119
  245,  4356.064
  246,  4371.713
  247,  4394.890
  248,  4412.097
  249,  4428.084
  250,  4444.254
  251,  4463.034
  252,  4479.685
  253,  4495.600
  254,  4511.888
  255,  4528.035
  256,  4544.097
  257,  4559.960
  258,  4575.919
  259,  4591.860
  260,  4607.827
  261,  4624.028
  262,  4642.172
  263,  4658.481
  264,  4676.100
  265,  4691.970
  266,  4708.537
  267,  4724.656
  268,  4740.581
  269,  4757.879
  270,  4774.302
  271,  4790.640
  272,  4806.363
  273,  4822.330
  274,  4838.335
  275,  4854.216
  276,  4869.838
  277,  4885.419
  278,  4902.662
  279,  4918.990
  280,  4934.687
  281,  4950.551
  282,  4966.351
  283,  4982.449
  284,  4999.076
  285,  5016.637
  286,  5033.854
  287,  5069.746
  288,  5086.579
  289,  5102.791
  290,  5118.531
  291,  5134.570
  292,  5150.221
  293,  5165.924
  294,  5184.497
  295,  5200.004
  296,  5215.665
  297,  5231.611
  298,  5247.550
  299,  5263.889
  300,  5280.191
  301,  5296.228
  302,  5311.781
  303,  5327.698
  304,  5344.277
  305,  5360.489
  306,  5376.537
  307,  5392.384
  308,  5408.279
  309,  5423.812
  310,  5439.569
  311,  5455.429
  312,  5471.057
  313,  5486.937
  314,  5502.451
  315,  5518.043
  316,  5533.592
  317,  5549.356
  318,  5565.116
  319,  5580.585
  320,  5596.141
  321,  5611.639
  322,  5627.209
  323,  5643.052
  324,  5659.013
  325,  5674.718
  326,  5690.320
  327,  5705.960
  328,  5721.852
  329,  5737.631
  330,  5753.383
  331,  5769.087
  332,  5784.730
  333,  5800.413
  334,  5815.796
  335,  5831.569
  336,  5846.994
  337,  5862.666
  338,  5878.185
  339,  5893.739
  340,  5909.518
  341,  5924.928
  342,  5940.309
  343,  5955.774
  344,  5971.499
  345,  5986.788
  346,  6002.143
  347,  6017.987
  348,  6033.518
  349,  6049.046
  350,  6064.706
  351,  6080.495
  352,  6096.055
  353,  6111.319
  354,  6126.759
  355,  6142.006
  356,  6157.545
  357,  6178.959
  358,  6194.108
  359,  6209.698
  360,  6225.010
  361,  6240.199
  362,  6255.398
  363,  6270.585
  364,  6286.564
  365,  6302.049
  366,  6318.744
  367,  6334.054
  368,  6349.539
  369,  6364.944
  370,  6380.835
  371,  6396.468
  372,  6411.864
  373,  6427.715
  374,  6442.849
  375,  6459.151
  376,  6474.790
  377,  6491.222
  378,  6506.929
  379,  6522.486
  380,  6537.795
  381,  6553.425
  382,  6569.025
  383,  6585.724
  384,  6600.954
  385,  6616.822
  386,  6632.287
  387,  6647.950
  388,  6663.144
  389,  6679.302
  390,  6694.788
  391,  6710.398
  392,  6726.113
  393,  6742.933
  394,  6758.249
  395,  6779.180
  396,  6804.005
  397,  6839.878
  398,  6855.939
  399,  6873.363
  400,  6888.349
  401,  6903.244
  402,  6919.060
  403,  6934.165
  404,  6982.046
  405,  6996.728
  406,  7011.619
  407,  7026.834
  408,  7054.801
  409,  7070.738
  410,  7085.838
  411,  7101.164
  412,  7115.819
  413,  7130.517
  414,  7145.099
  415,  7160.033
  416,  7178.238
  417,  7193.395
  418,  7208.511
  419,  7223.630
  420,  7238.425

Either way. I'm done with you. If you dont like SLI. I could give a rip. Don't buy it again. But to say people shouldnt because you are bothered by a largely subjective intermittent change in frame syncing. Its not anymore fair than me saying people should appreciate SLI for being flawless. Because it's not. Regardless, I like SLI. And I know alot of other people who do too. And no, They aren't ignorant people. And frankly I don't care that you care if I think SLI is worth while. If people in this thread are only interested in seeing a negative viewpoint on SLI and Crossfire. I'm sorry. But I dont have one and I'll gladly offer my opinion and own data if I damn well please. Until beyond3d decides that my input isn't welcome here.

 

[annihilator]

If a MS delay corresponding to momentary FPS surges from 45FPS to 70FPS all the time (and this is from the 2-way UT3 example YOU gave. In the 3dmark example I gave, which you don't even talk a word about, it fluctuates from 20 to 50) are "largely unnoticeable" to you then I suggest you don't bother with SLI anyway because a FPS difference between 45 and 70 isn't largely noticeable for your eyes, and you could largely do without SLI.

 

[ChrisRay]

I'm sorry. But I'm not interested in playing games at low details and staggeringly low frame rates. 1 frame taking 100 MS to render is largely unacceptable to my eyes. And I'd rather have a slightly intermittent AFR synchronization than run a game at 10 FPS. Which brings to my last point. Perfectly Synced frames will cause SLI to fail ((15-30-45-60-75-90-105-120-135-150-165.)) I've seen it happen and all it does is cause the games performance to drop radically over time. And cause frames to dip out and grow worse as time progresses. The frame MS delay just gets larger and larger. Whats so ridiculous about all of this is inconsistent MS delays occur with single GPUs too.

P.S. I dont care about 3dmark06.

 

[compres]

I'm sorry. But I'm not interested in playing games at low details and staggeringly low frame rates. 1 frame taking 100 MS to render is largely unacceptable to my eyes. And I'd rather have a slightly intermittent AFR synchronization than run a game at 10 FPS. Which brings to my last point. Perfectly Synced frames will cause SLI to fail ((15-30-45-60-75-90-105-120-135-150-165.)) I've seen it happen and all it does is cause the games performance to drop radically over time. And cause frames to dip out and grow worse as time progresses. The frame MS delay just gets larger and larger. Whats so ridiculous about all of this is inconsistent MS delays occur with single GPUs too.

P.S. I dont care about 3dmark06.

That is very very true. While reading I am just wondering how much worse it is on multi GPUs.

 

[MfA]

What exactly is this mythical load balancing I keep hearing about? (For AFR that is, it's easy to see how it works with SFR ... but for AFR it makes no sense at all to even think in terms of load balancing AFAICS.)

 

[SirPauly]

What would be the correct wording to use then so there may be more clarity on the issue?

 

[MfA]

First tell me what it does.

 

[SirPauly]

I'm not an expert --- just looking for the correct wording since load balancing with AFR seems to be mythical from your point-of-view or experience. Was hoping you would be kind enough to at least answer.

Personally always used the term "load balancing" based on smoothness issues at times with AFR and would like to know if I am way off base and very ignorant -- and if so, sorry, but willing to listen and learn.