Detonator 52.14 Test & 52.10 Re-Test

October 8, 2003 / by Leonidas / page 6 of 7

   Shader Performance

On to a much more pleasant issue: shader performance of the Detonators 52.10 and 52.14. It was measured again using ShaderMark 2.0 with the setting "Shader 2.a". The following benchmarks were conducted using the same test system as before and without any quality enhancing features like anistotropic filtering or anti aliasing (omitted shaders couldn't be displayed correctly):

	45.23	52.10	52.14	45.23 (partial accuracy)	52.10 (partial accuracy)	52.14 (partial accuracy)
Shader 1	-	-	-	-	-	-
Shader 2	94 fps	118 fps	118 fps	108 fps	138 fps	138 fps
Shader 3	60 fps	76 fps	76 fps	81 fps	98 fps	98 fps
Shader 4	-	-	-	-	-	-
Shader 5	44 fps	65 fps	65 fps	64 fps	79 fps	79 fps
Shader 6	61 fps	73 fps	73 fps	88 fps	109 fps	109 fps
Shader 7	44 fps	68 fps	68 fps	69 fps	94 fps	94 fps
Shader 8	-	-	-	-	-	-
Shader 9	30 fps	36 fps	36 fps	46 fps	54 fps	55 fps
Shader 10	116 fps	145 fps	145 fps	117 fps	144 fps	144 fps
Shader 11	100 fps	117 fps	117 fps	108 fps	125 fps	125 fps
Shader 12	45 fps	61 fps	61 fps	66 fps	79 fps	80 fps
Shader 13	26 fps	36 fps	36 fps	45 fps	57 fps	57 fps
Shader 14	36 fps	47 fps	47 fps	59 fps	71 fps	71 fps
Shader 15	46 fps	46 fps	46 fps	66 fps	76 fps	76 fps
Shader 16	-	37 fps	37 fps	-	53 fps	54 fps
Shader 17	4 fps	5 fps	5 fps	6 fps	9 fps	9 fps
Shader 18	16 fps	28 fps	28 fps	37 fps	49 fps	49 fps
Shader 19	-	-	-	-	-	-
Shader 20	-	-	-	-	-	-
Shader 21	-	-	-	-	-	-
Shader 22	-	-	-	-	-	-
Shader 23	-	-	-	-	-	-

Of course we cannot be absolutely sure that the faked trilinear filtering of the Detonators 52.10 and 52.14 doesn't contribute a small fraction to the measured performance speed-up. But we think it's quite unlikely that filtering performance plays such a significant role running these shaders, so it could explain the drastical performance enhancements.

We conclude at the moment that nVidia really improved shader performance with the 52.10/14 drivers. In the measurement with full accuracy (32bits per color channel), Detonator 52.10 achieved a whopping 31.5% speed-up and with partial accuracy (16bits per color channel) at least 25% more performance. The scores of Detonator 52.14 are nearly identical to those of Detonator 52.10, only in the measurements with partial accurancy the Detonator 52.14 is in three tests 1fps faster.

It has yet to be proven whether nVidia can turn this theoretical advantage in more performance in real applications, but because nVidia falls quite far behind Ati in benchmarks which rely on shader performance, those 31.5% won't be enough to get on par with ATi entirely. However, the distance to ATi shrinked a bit for nVidia with the plus in shader performance introduced with the 52.10/14 drivers.

   GeForce4 Ti4200 Test

Something is still missing: a test with a video card from the previous generation. So we inserted an AOpen GeForce4 Ti4200 with 64MB in the test system and confronted it with the Detonators 52.10 and 52.14. With the D3D AF-Tester, we discovered that Detonator 52.14 is indeed able to do correct trilinear filtering (MouseOver effect using Javascript, alternative: click opens both screenshots):

Set in the control panel as well as in Application mode, the 52.10 and 52.14 drivers are filtering exactly like we want 'em to - without any optimizations - on the GeForce4 Ti4200. A test with Unreal Tournament 2003 showed no sign of an application specific optimization, too. So all conclusions so far only apply to video cards of the GeForceFX series and not of the GeForce4 series.

We won't conduct a benchmark with the GeForce4 Ti4200 here because graphics chip developers usually only optimize for their current chips and never for older ones. Furthermore, the performance enhancements measured using the GeForceFX 5900 are likely a result of the "optimizations" found in the Detonators 52.10 and 52.14, so a GeForce4 Ti which doesn't utilize those "optimizations" also shouldn't experience any performance increase.