:strip_icc():strip_exif()/u/55093/UPpharoah.jpg?f=community)
Voordat je je reactie geeft, mag je dit volgenden artikel even doorspitten. Het komt van http://www.extremetech.com en behalve de laatste 3.30 resultaten gaan ze ook even in op de mogelijkheden en onmogelijkheden van benchmarks.
UPDATED ATI TEST RESULTS
In our original story down below, we presented extensive test data that showed how both ATI and nVidia were losing some performance in the newest 3.30 version of 3DMark03, which has implemented countermeasures to defeat app detection by 3D GPU drivers. Due to a variation in our test methodology, the drop seen in our initial tests for ATI appears to be have been unduly affected by the location where render state settings are made (in the application versus in the driver).
Our original test results misstated ATI's performance drop with 3DMark03 version 3.30, and we'll provide correct, updated results that set ATI in a much better light, with far less performance hit moving to version 3.30. However, ATI still detected 3DMark03 within their driver, and substituted a "better optimized" shader, and that is still benchmark cheating in our book, even if the overall performance boost was modest.
In other words, the results we initially reported were more pronounced than they should have been for ATI– holding all else equal. Based on our retesting, we still saw that ATI takes a small hit on Game Test 4 in version 3.30, which has to do with a driver "optimization" that ATI has already publicly stated will be removed from its next Catalyst driver.
We'll explain how these issues arose, and present the updated test results for ATI.
We went back and tested on our same Intel reference test system, which has the following configuration:
Intel Pentium 4 3GHz
Intel 875PZB motherboard using Intel 875P chipset
512MB 400MHz DDR SDRAM
Sound Blaster Audigy Gamer
3com NIC
40GB ATA-100 EIDE hard drive
Toshiba DVD-ROM
KDS Avitron 21" monitor
Fresh install of Windows XP Pro SP1 with latest system updates as of 4/25/03
DirectX 9.0a
We tested at 1280x960x32 and 1600x1200x32, and with both 4X AA and 8X AF enabled at both resolutions.
Who's Driving Here?
The first issue we hit upon has to do with where anisotropic filtering settings are made. We took a long, hard look in the past at both ATI's and nVidia's adaptive implementations of AF, since both companies apply AF on an as-needed basis. Anisotropic filtering is used to address a specific kind of texture artifact that occurs when a 3D surface is sloped relative to the view camera.
Performance differences can arise depending on whether an application sets the AF parameters, or whether these are forced on at a driver level. There are additional parameters that applications can specify regarding how AF is applied to a scene, starting with which parts of the scene will have AF applied and which will not.
When AF is forced onto an unsuspecting app at a driver level, the driver attempts to make best-guess assumptions about how to apply AF, and not significantly degrade frame-rate performance.
In the case of 3DMark03, when AF is enabled from the benchmark itself, additional AF work has to be done during the four game test scenes, and this extra work induces roughly an 8% performance hit for ATI on its overall 3DMark03 score, compared to when the AF state is set at the driver level (by a setting in the graphics control panel).
Both nVidia and ATI allow their different AF settings to be set in their driver control panels. nVidia has a single panel that governs both Direct3D and OpenGL, whereas ATI has a panel for each of the two APIs, making for more "knobs and levers," but also allowing more granularity of these settings.
During our original testing of 3DMark03 version 3.30, ATI cards were set to use AF settings specified by 3DMark03, not the ATI driver's control panel. This created a much bigger drop in performance than if we had used the ATI driver's AF settings only. For the nVidia results, we used the driver control panel settings for both version 3.20 and 3.30 of 3DMark03, and not 3DMark03's AF settings, so our nVidia results will stand as previously tested. With the new results, ATI is using driver setting for AF in all cases, so it's more of an apples-to-apples comparison of ATI versus nVidia.
Here we present the data for both versions 3.20 and 3.30 of 3DMark03, with the latter version having anti-detection measures built into it to defeat driver detection/optimizations/funny business.
Note in our chart that "APP" refers to when anisotropic filtering is implemented via the 3DMark2003 benchmark, while "DRV" refers to setting AF via the driver before running the benchmarks.
Overall 3DMark03 Score
Version 3.20 Version 3.30 %drop v. 3.30 vs. v. 3.20
R9800 12x9 - DRV 2214 2150 2.9%
R9800 12x9 - APP 2028 1975 2.6%
R9800 16x12 - DRV 1613 1552 3.8%
R9800 16x12 - APP 1477 1430 3.2%
As you can see from the above chart, ATI's overall performance drops less than 3% at 1280x960 -- within the margin of error for the test – when AF is applied consistently using either the driver method or the application method. This occurs when running both versions 3.20 and 3.30 of 3DMark03. (See new ATI versus nVidia results below).
At 16x12, the difference grows a little more marked – almost 4% at the driver level, which is starting to become significant, but still not tremendously different.
App-level versus Driver-level AF Settings
This next chart shows the differences in overall 3DMark03 score if we set AF on from the application in one run, and then from the driver in the other. We present both the percentage decrease by using APP vs. DRV, and the percentage increase of using DRV vs. APP, two sides of the same coin, really. Letting 3DMark2003 set the AF results in about 9% drop in performance consistently whether you test with version 3.30 or version 3.20 of 3DMark2003.
Ver. 3.20 %Drop using APP Ver. 3.20 %Gain Using DRV Ver. 3.30 %Drop using APP Ver. 3.30 %Gain Using DRV
R9800 12x9 8.4% 9.2% 8.1% 8.9%
R9800 16x12 8.4% 9.2% 7.9% 8.5%
The 9% gain ATI gets at both resolutions by using its own driver-level AF setting pulls ATI even with the GeForceFX 5900 Ultra, which we also tested using that GPU's driver-level AF settings. When we initially (and erroneously) compared ATI's application-set scores versus nVidia's driver-set scores on version 3.30, nVidia was ahead by about 10%.
Differences Still Seen in Game Test 4
3DMark2003 game test 4 was the only test that showed significant performance differences for ATI, whether testing was done consistently at the application or the driver level. Here are the frame rate results from Game Test 4 – Mother Nature, the DX9/Pixel Shaders 2.0 test.
Version 3.20 Version 3.30 %drop v. 3.30 vs. v. 3.20
R9800 12x9 - DRV 20.5 18.8 8.3%
R9800 12x9 - APP 17.9 15.8 11.7%
R9800 16x12 - DRV 15.8 14.2 10.1%
R9800 16x12 - APP 13.8 12.6 8.7%
What you can see here is that even the biggest percentage drop going from version 3.20 to version 3.30, 11.7%, turns out to be all of 2.1fps, not a significant actual amount of performance loss. These deltas indicate how frame rate data, taken as percentages, can obscure, or in this case, artificially augment what's really going on.
When you vary how AF is set in Game Test 4, from application to driver and back, the differences are even more pronounced, as you can see here.
Ver. 3.20 %Drop using APP Ver. 3.20 %Gain Using DRV Ver. 3.30 %Drop using APP Ver. 3.30 %Gain Using DRV
R9800 12x9 - APP 12.7% 14.5% 16.0% 19.0%
R9800 16x12 - APP 12.7% 14.5% 11.3% 12.7%
The nVidia/ATI Matchup
The saucier question is, do the results seen in version 3.30 of 3DMark03 affect how nVidia and ATI's match up on this benchmark? And the answer is, yes, except that it now becomes that much more difficult to declare an out and out winner.
nVidia holds a noteworthy lead at both test resolutions when running on version 3.20, but that's with cheats in place, so the results don't count. nVidia's lead gets all but obliterated in version 3.30 when we also use ATI's driver-level AF settings.
Here are the raw scores with AF set at the driver level on both products.
1280x960x32 V 3.20 Version 3.30 %drop v. 3.30 vs. v. 3.20
ATI R9800 - DRV 2214 2150 2.9%
nVidia GFFX 5900 Ultra - DRV 2623 2182 16.8%
1600x1200x32 Version 3.20 Version 3.30 %drop V/V
ATI R9800 - DRV 1613 1552 3.8%
nVidia GFFX 5900 Ultra - DRV 1743 1567 10.1%
Below we list the winners by percentage. As you can see, nVidia is ahead of ATI by a negligible 1% on both tests.
Winner, version 3.20 Winner, Version 3.30
1280x960x32 nVidia, by 18.5% nVidia, by 1.5%
1600x1200x32 nVidia, by 8.1% nVidia, by 1%
Because Game Test 4 is where we saw the biggest deltas, here are the comparative ATI vs nVidia results for your perusing pleasure:
GAME TEST 4
1280x960x32 Version 3.20 Version 3.30
ATI R9800 - DRV 20.5 18.8
nVidia GFFX 5900 Ultra - DRV 17.9 9.9
1600x1200x32
ATI R9800 - DRV 15.8 14.2
nVidia GFFX 5900 Ultra - DRV 11.7 7.3
As you can see, ATI does take a slight hit, around 8% and 10% at 12x9 and 16x12, respectively. But again, consider that the actual frame rate delta is never more than 1.7fps, a minute amount.
nVidia however suffers larger hits, and at 12x9, sees its score nearly cut in half. It's also interesting to note that nVidia delivers only half of the performance of ATI's 9800 Ultra, indicating some real shader-based performance differences between the parts here – and perhaps a motive for nVidia to artificially pump up its scores. Still, this is all of one data point. We still don't have any games that use DX9's Pixel Shaders 2.0 standard, and until we do, the question of which GPU has superior performance in this area remains open.
Too Many Knobs and Levers...
We had originally reported in the story below that ATI's performance on Game Test 4 dropped by around 20% at both test resolutions. This seemingly large drop amounted to a mere 3fps, but the delta turns out to be about half that amount, as we showed above.
So what happened? We suspect that in preparing our version 3.30 numbers, we went into ATI's driver control panel, and set the correct AF and AA parameters, 8X and 4X, respectively. But upon launching the application, we proceeded to also set those same parameters at an application level, which over-rode the driver-level settings, and had ATI doing the work specified by the app, rather than letting the driver handle AF filtering decisions.
We've added an additional step to our testing methodology, to verify AA and AF settings at a driver level only, and leave the AF and AA settings at their default state in the benchmark.
Did ATI "cheat" on 3DMark2003?
Version 3.30 of 3DMark03 not only defeated the DetonatorFX driver's app detection mechanisms, but it also managed to uncover an ATI optimization that landed the company in a bit of hot water.
ATI's optimization involves using a more optimized shader program in Game Test 4 for the water shader program, but according to ATI, this water shader renders "…the scene exactly as intended by Futuremark, in full-precision floating point. Our shaders are mathematically and functionally identical to Futuremark's, and there are no visual artifacts; we simply shuffle instructions to take advantage of our architecture….However, we recognize that these can be used by some people to call into question the legitimacy of benchmark results, and so we are removing them from our driver as soon as is physically possible. We expect them to be gone by the next release of CATALYST."
The line between a bona fide optimization and a cheat is sometimes not well-defined, but in our estimation, ATI's optimization is suspect, and at the end of the day, it bought them all of a 9% increase in performance, or a barely mention-worthy 1.7fps.
This small gain leads us to question why ATI would implement this optimization in the first place, since the risk of being perceived as cheating would far outweigh a less than 2fps performance increase. However, ATI has been forthcoming in its public statements about the nature of this optimization, and has stated that because of the potential perception of being considered a cheat (and probably because it didn't buy them much performance anyway), that this optimization will be pulled from the next Catalyst driver release.
Meanwhile, nVidia's "optimizations" are far more aggressive, and we believe they go beyond what can be considered reasonable driver optimization, and into the realm of cheating. nVidia has thus far been unable, or perhaps unwilling, to address these issues, since it apparently still refuses to rejoin FutureMark's beta program to gain access to the developer version of 3DMark03, and the company maintains that what ExtremeTech uncovered are driver bugs, and nothing more.
Lessons Learned
It's clear that applications of all stripes need to factor in AF and AA into their settings control panels, and that game developers need to intelligently implement how these two features get used in their games, rather than leaving GPU drivers to make educated guesses as to what will be best. Serious Sam SE does allow for tweaking of AF settings in its advanced rendering options panel, although this option is buried three screens down in those panels.
It also raises once again the issue of subjective versus objective evaluation of 3D performance. Image quality is very important, and anything a GPU can do to palpably improve image quality and the overall gaming experience is obviously a good thing.
But dynamic image quality optimizations can make benchmarking nearly impossible, since one cannot know what exactly a GPU is doing. It's for this reason for example that we always test nVidia-based GPUs with the 3D Quality and Performance slider set at its "Application" position. Yes, nVidia can do intelligent things to dynamically improve image quality and performance, but how much of this is being done at any given time is impossible to say.
Then there is the matter of less sophisticated users ignoring these somewhat esoteric features. This is where nVidia's over-arching "3D goodness" slider is helpful, since it allows someone to simply dial up the 3D love without having to know exactly what it does.
The problem here is one of competing objectives: performance sliders such as this help end users, but can drive analysts nuts. The trick, it would seem, is to serve the sometimes-conflicting needs of end-users and analysts alike, and simultaneously improve the 3D user experience of the former without making the job of the latter impossible. One simple way to make this workable is to be sure that whatever consumer-friendly "3D goodness sliders" are added to driver panels, that they can be turned off/disabled completely, forcing the GPU to render the scene as the application has specified. In this way, we can look at "absolute" performance, and then see how much performance scales by using the 3D goodness slider, and at the same time verify that image quality isn't being adversely affected in a palpable way.
The actual performance difference seen in our new test results is fairly insignificant in terms of actual frame rate, but for the purposes of consistent methodology, we needed to revisit this matter and make sure we provided you with accurate information. To avoid such snafus in the future, we'll be using driver-level settings for AF on both GPUs at all times, and forgo using 3DMark's app-level AF settings. This methodology better reflects the current landscape today, where the majority of apps don't understand AF, and don't have settings for it in their options panels.
UPDATED ATI TEST RESULTS
In our original story down below, we presented extensive test data that showed how both ATI and nVidia were losing some performance in the newest 3.30 version of 3DMark03, which has implemented countermeasures to defeat app detection by 3D GPU drivers. Due to a variation in our test methodology, the drop seen in our initial tests for ATI appears to be have been unduly affected by the location where render state settings are made (in the application versus in the driver).
Our original test results misstated ATI's performance drop with 3DMark03 version 3.30, and we'll provide correct, updated results that set ATI in a much better light, with far less performance hit moving to version 3.30. However, ATI still detected 3DMark03 within their driver, and substituted a "better optimized" shader, and that is still benchmark cheating in our book, even if the overall performance boost was modest.
In other words, the results we initially reported were more pronounced than they should have been for ATI– holding all else equal. Based on our retesting, we still saw that ATI takes a small hit on Game Test 4 in version 3.30, which has to do with a driver "optimization" that ATI has already publicly stated will be removed from its next Catalyst driver.
We'll explain how these issues arose, and present the updated test results for ATI.
We went back and tested on our same Intel reference test system, which has the following configuration:
Intel Pentium 4 3GHz
Intel 875PZB motherboard using Intel 875P chipset
512MB 400MHz DDR SDRAM
Sound Blaster Audigy Gamer
3com NIC
40GB ATA-100 EIDE hard drive
Toshiba DVD-ROM
KDS Avitron 21" monitor
Fresh install of Windows XP Pro SP1 with latest system updates as of 4/25/03
DirectX 9.0a
We tested at 1280x960x32 and 1600x1200x32, and with both 4X AA and 8X AF enabled at both resolutions.
Who's Driving Here?
The first issue we hit upon has to do with where anisotropic filtering settings are made. We took a long, hard look in the past at both ATI's and nVidia's adaptive implementations of AF, since both companies apply AF on an as-needed basis. Anisotropic filtering is used to address a specific kind of texture artifact that occurs when a 3D surface is sloped relative to the view camera.
Performance differences can arise depending on whether an application sets the AF parameters, or whether these are forced on at a driver level. There are additional parameters that applications can specify regarding how AF is applied to a scene, starting with which parts of the scene will have AF applied and which will not.
When AF is forced onto an unsuspecting app at a driver level, the driver attempts to make best-guess assumptions about how to apply AF, and not significantly degrade frame-rate performance.
In the case of 3DMark03, when AF is enabled from the benchmark itself, additional AF work has to be done during the four game test scenes, and this extra work induces roughly an 8% performance hit for ATI on its overall 3DMark03 score, compared to when the AF state is set at the driver level (by a setting in the graphics control panel).
Both nVidia and ATI allow their different AF settings to be set in their driver control panels. nVidia has a single panel that governs both Direct3D and OpenGL, whereas ATI has a panel for each of the two APIs, making for more "knobs and levers," but also allowing more granularity of these settings.
During our original testing of 3DMark03 version 3.30, ATI cards were set to use AF settings specified by 3DMark03, not the ATI driver's control panel. This created a much bigger drop in performance than if we had used the ATI driver's AF settings only. For the nVidia results, we used the driver control panel settings for both version 3.20 and 3.30 of 3DMark03, and not 3DMark03's AF settings, so our nVidia results will stand as previously tested. With the new results, ATI is using driver setting for AF in all cases, so it's more of an apples-to-apples comparison of ATI versus nVidia.
Here we present the data for both versions 3.20 and 3.30 of 3DMark03, with the latter version having anti-detection measures built into it to defeat driver detection/optimizations/funny business.
Note in our chart that "APP" refers to when anisotropic filtering is implemented via the 3DMark2003 benchmark, while "DRV" refers to setting AF via the driver before running the benchmarks.
Overall 3DMark03 Score
Version 3.20 Version 3.30 %drop v. 3.30 vs. v. 3.20
R9800 12x9 - DRV 2214 2150 2.9%
R9800 12x9 - APP 2028 1975 2.6%
R9800 16x12 - DRV 1613 1552 3.8%
R9800 16x12 - APP 1477 1430 3.2%
As you can see from the above chart, ATI's overall performance drops less than 3% at 1280x960 -- within the margin of error for the test – when AF is applied consistently using either the driver method or the application method. This occurs when running both versions 3.20 and 3.30 of 3DMark03. (See new ATI versus nVidia results below).
At 16x12, the difference grows a little more marked – almost 4% at the driver level, which is starting to become significant, but still not tremendously different.
App-level versus Driver-level AF Settings
This next chart shows the differences in overall 3DMark03 score if we set AF on from the application in one run, and then from the driver in the other. We present both the percentage decrease by using APP vs. DRV, and the percentage increase of using DRV vs. APP, two sides of the same coin, really. Letting 3DMark2003 set the AF results in about 9% drop in performance consistently whether you test with version 3.30 or version 3.20 of 3DMark2003.
Ver. 3.20 %Drop using APP Ver. 3.20 %Gain Using DRV Ver. 3.30 %Drop using APP Ver. 3.30 %Gain Using DRV
R9800 12x9 8.4% 9.2% 8.1% 8.9%
R9800 16x12 8.4% 9.2% 7.9% 8.5%
The 9% gain ATI gets at both resolutions by using its own driver-level AF setting pulls ATI even with the GeForceFX 5900 Ultra, which we also tested using that GPU's driver-level AF settings. When we initially (and erroneously) compared ATI's application-set scores versus nVidia's driver-set scores on version 3.30, nVidia was ahead by about 10%.
Differences Still Seen in Game Test 4
3DMark2003 game test 4 was the only test that showed significant performance differences for ATI, whether testing was done consistently at the application or the driver level. Here are the frame rate results from Game Test 4 – Mother Nature, the DX9/Pixel Shaders 2.0 test.
Version 3.20 Version 3.30 %drop v. 3.30 vs. v. 3.20
R9800 12x9 - DRV 20.5 18.8 8.3%
R9800 12x9 - APP 17.9 15.8 11.7%
R9800 16x12 - DRV 15.8 14.2 10.1%
R9800 16x12 - APP 13.8 12.6 8.7%
What you can see here is that even the biggest percentage drop going from version 3.20 to version 3.30, 11.7%, turns out to be all of 2.1fps, not a significant actual amount of performance loss. These deltas indicate how frame rate data, taken as percentages, can obscure, or in this case, artificially augment what's really going on.
When you vary how AF is set in Game Test 4, from application to driver and back, the differences are even more pronounced, as you can see here.
Ver. 3.20 %Drop using APP Ver. 3.20 %Gain Using DRV Ver. 3.30 %Drop using APP Ver. 3.30 %Gain Using DRV
R9800 12x9 - APP 12.7% 14.5% 16.0% 19.0%
R9800 16x12 - APP 12.7% 14.5% 11.3% 12.7%
The nVidia/ATI Matchup
The saucier question is, do the results seen in version 3.30 of 3DMark03 affect how nVidia and ATI's match up on this benchmark? And the answer is, yes, except that it now becomes that much more difficult to declare an out and out winner.
nVidia holds a noteworthy lead at both test resolutions when running on version 3.20, but that's with cheats in place, so the results don't count. nVidia's lead gets all but obliterated in version 3.30 when we also use ATI's driver-level AF settings.
Here are the raw scores with AF set at the driver level on both products.
1280x960x32 V 3.20 Version 3.30 %drop v. 3.30 vs. v. 3.20
ATI R9800 - DRV 2214 2150 2.9%
nVidia GFFX 5900 Ultra - DRV 2623 2182 16.8%
1600x1200x32 Version 3.20 Version 3.30 %drop V/V
ATI R9800 - DRV 1613 1552 3.8%
nVidia GFFX 5900 Ultra - DRV 1743 1567 10.1%
Below we list the winners by percentage. As you can see, nVidia is ahead of ATI by a negligible 1% on both tests.
Winner, version 3.20 Winner, Version 3.30
1280x960x32 nVidia, by 18.5% nVidia, by 1.5%
1600x1200x32 nVidia, by 8.1% nVidia, by 1%
Because Game Test 4 is where we saw the biggest deltas, here are the comparative ATI vs nVidia results for your perusing pleasure:
GAME TEST 4
1280x960x32 Version 3.20 Version 3.30
ATI R9800 - DRV 20.5 18.8
nVidia GFFX 5900 Ultra - DRV 17.9 9.9
1600x1200x32
ATI R9800 - DRV 15.8 14.2
nVidia GFFX 5900 Ultra - DRV 11.7 7.3
As you can see, ATI does take a slight hit, around 8% and 10% at 12x9 and 16x12, respectively. But again, consider that the actual frame rate delta is never more than 1.7fps, a minute amount.
nVidia however suffers larger hits, and at 12x9, sees its score nearly cut in half. It's also interesting to note that nVidia delivers only half of the performance of ATI's 9800 Ultra, indicating some real shader-based performance differences between the parts here – and perhaps a motive for nVidia to artificially pump up its scores. Still, this is all of one data point. We still don't have any games that use DX9's Pixel Shaders 2.0 standard, and until we do, the question of which GPU has superior performance in this area remains open.
Too Many Knobs and Levers...
We had originally reported in the story below that ATI's performance on Game Test 4 dropped by around 20% at both test resolutions. This seemingly large drop amounted to a mere 3fps, but the delta turns out to be about half that amount, as we showed above.
So what happened? We suspect that in preparing our version 3.30 numbers, we went into ATI's driver control panel, and set the correct AF and AA parameters, 8X and 4X, respectively. But upon launching the application, we proceeded to also set those same parameters at an application level, which over-rode the driver-level settings, and had ATI doing the work specified by the app, rather than letting the driver handle AF filtering decisions.
We've added an additional step to our testing methodology, to verify AA and AF settings at a driver level only, and leave the AF and AA settings at their default state in the benchmark.
Did ATI "cheat" on 3DMark2003?
Version 3.30 of 3DMark03 not only defeated the DetonatorFX driver's app detection mechanisms, but it also managed to uncover an ATI optimization that landed the company in a bit of hot water.
ATI's optimization involves using a more optimized shader program in Game Test 4 for the water shader program, but according to ATI, this water shader renders "…the scene exactly as intended by Futuremark, in full-precision floating point. Our shaders are mathematically and functionally identical to Futuremark's, and there are no visual artifacts; we simply shuffle instructions to take advantage of our architecture….However, we recognize that these can be used by some people to call into question the legitimacy of benchmark results, and so we are removing them from our driver as soon as is physically possible. We expect them to be gone by the next release of CATALYST."
The line between a bona fide optimization and a cheat is sometimes not well-defined, but in our estimation, ATI's optimization is suspect, and at the end of the day, it bought them all of a 9% increase in performance, or a barely mention-worthy 1.7fps.
This small gain leads us to question why ATI would implement this optimization in the first place, since the risk of being perceived as cheating would far outweigh a less than 2fps performance increase. However, ATI has been forthcoming in its public statements about the nature of this optimization, and has stated that because of the potential perception of being considered a cheat (and probably because it didn't buy them much performance anyway), that this optimization will be pulled from the next Catalyst driver release.
Meanwhile, nVidia's "optimizations" are far more aggressive, and we believe they go beyond what can be considered reasonable driver optimization, and into the realm of cheating. nVidia has thus far been unable, or perhaps unwilling, to address these issues, since it apparently still refuses to rejoin FutureMark's beta program to gain access to the developer version of 3DMark03, and the company maintains that what ExtremeTech uncovered are driver bugs, and nothing more.
Lessons Learned
It's clear that applications of all stripes need to factor in AF and AA into their settings control panels, and that game developers need to intelligently implement how these two features get used in their games, rather than leaving GPU drivers to make educated guesses as to what will be best. Serious Sam SE does allow for tweaking of AF settings in its advanced rendering options panel, although this option is buried three screens down in those panels.
It also raises once again the issue of subjective versus objective evaluation of 3D performance. Image quality is very important, and anything a GPU can do to palpably improve image quality and the overall gaming experience is obviously a good thing.
But dynamic image quality optimizations can make benchmarking nearly impossible, since one cannot know what exactly a GPU is doing. It's for this reason for example that we always test nVidia-based GPUs with the 3D Quality and Performance slider set at its "Application" position. Yes, nVidia can do intelligent things to dynamically improve image quality and performance, but how much of this is being done at any given time is impossible to say.
Then there is the matter of less sophisticated users ignoring these somewhat esoteric features. This is where nVidia's over-arching "3D goodness" slider is helpful, since it allows someone to simply dial up the 3D love without having to know exactly what it does.
The problem here is one of competing objectives: performance sliders such as this help end users, but can drive analysts nuts. The trick, it would seem, is to serve the sometimes-conflicting needs of end-users and analysts alike, and simultaneously improve the 3D user experience of the former without making the job of the latter impossible. One simple way to make this workable is to be sure that whatever consumer-friendly "3D goodness sliders" are added to driver panels, that they can be turned off/disabled completely, forcing the GPU to render the scene as the application has specified. In this way, we can look at "absolute" performance, and then see how much performance scales by using the 3D goodness slider, and at the same time verify that image quality isn't being adversely affected in a palpable way.
The actual performance difference seen in our new test results is fairly insignificant in terms of actual frame rate, but for the purposes of consistent methodology, we needed to revisit this matter and make sure we provided you with accurate information. To avoid such snafus in the future, we'll be using driver-level settings for AF on both GPUs at all times, and forgo using 3DMark's app-level AF settings. This methodology better reflects the current landscape today, where the majority of apps don't understand AF, and don't have settings for it in their options panels.
Voor upgrade, complete zelfbouw, herstellen van windows/software klik hier!
:strip_exif()/u/80778/wacomm-test-I60.gif?f=community)
:strip_icc():strip_exif()/u/23850/aquarius.jpg?f=community)
Lang leve de concurrentie 
:strip_icc():strip_exif()/u/63098/crop651ad86823c45.jpg?f=community)
u said it all.
)
/u/61371/crop65e9c74564967_cropped.png?f=community)
:strip_icc():strip_exif()/u/30256/right_of_way_065.jpg?f=community)
/u/76318/crop5e61009512718.png?f=community)
/u/18014/icon2.png?f=community)
Dit topic is gesloten.