A new OpenGL BenchMark ?
- Thread starter nelg
- Start date
Current @1600x1200x32: 392fps (after repeat: 403fps)
Pixelshader : 266fps (after repeat 271fps)
Pixelshader 2.0 Low Prec: 37 fps (same after repeat)
Pixelshader 2.0 High Prec: 37fps (same after repeat)
Image Resolution in Windows was 1280x1024x32, the image was rendered by FX5800u with ForceWare 53.03 drivers.
Pixelshader : 266fps (after repeat 271fps)
Pixelshader 2.0 Low Prec: 37 fps (same after repeat)
Pixelshader 2.0 High Prec: 37fps (same after repeat)
Image Resolution in Windows was 1280x1024x32, the image was rendered by FX5800u with ForceWare 53.03 drivers.
Mendel said:I bet this disables the Unified Shader Compiler too
I don't know what you mean, but
1.) FP16 is forced by the driver on NV<35 chipsets, so there's no speed difference.
2.) Due to FP16 and low number of FP units, NV<35 chips are operating at they theoretical speed limit.
(therefore)
3.) The shader compiler enhancements benefited the NV>=35 chipsets only.
Those have 3x FP power with the same register limitation problem -> they are extremely sensitive to instruction scheduling.
@Hyp-X:
I don't know about DirectX, but in OpenGL using 16 or 32 bit floats results in the same speed and different precision(I was writing results to a floating-point buffer and reading them back).
The advantages about using 16 bit floats on GFX is only about register issue, but it doesn't matter since unified compiler.
I don't know about DirectX, but in OpenGL using 16 or 32 bit floats results in the same speed and different precision(I was writing results to a floating-point buffer and reading them back).
The advantages about using 16 bit floats on GFX is only about register issue, but it doesn't matter since unified compiler.
Ahmed Bahaa
Newcomer
Hi , i am Ahmed Bahaa , the Author of the program
i was checking my log file when i found a traffic coming from
this forum (which of course is most welcome) , as a matter
of fact i always relied on constructive discussion forums (just like
this one) to help me enhance my software as users are always able
to point me directions that i wasn't able to see in the first place.
Regarding the subject that the high precision pixelshader test result
is the same like the low precision , for ATI this is reasonable because
it has only one presision mode (24 bit) , so for high precision the
ATI card will execute 24 bits of precision , and for low precision it will
also execute 24 bits but will clamp the results afterwords to produce
16 bits of precision.
For Nvidia cards things are different , they support 12 bit , 16 bit , 32 bit
level of precision , so there must be a difference despite of if the
shader compiler used or not , because the shader compiler only change
the ordering of the execution of the program instruction but never changes
the precision , (or at least that's what nvidia says) , i wasn't able to test
the benchmark in GeforceFX cards because i only have a Radeon 9700 pro
,Radean 8500,Geforce4 TI4600 and Geforce3 TI500 at the moment ,
when i received reports that the GeforceFX is giving similer results i had
a closer look at the shader code and discoverd a bug in the high precision
pixelshader test led to it is always executing as low precision , i deeply
apologize for this mistake as the fact that i don't have GeforceFX and the
results should be identical on the ATI card of mine prevented me from
detecting this error before Release , now i corrected the shader code
along with some general modifications for the program based on the
very helpfull user opinions , and posted a new Release 1.1
(which you can download it at http://www.aureliosoft.com/download.html)
ofcourse the results could be the same after the modification as well
if any GeforceFX owner can post results here it will be of a great help
Also if anyone has comments or suggestins on the benchmark or the
program itself they will be wellcomed.
I want to thank you all for spending your valuable time for this discussion
it was realy of a great help , and i wish we could share more constructive thoughts ,
Ahmed Bahaa
i was checking my log file when i found a traffic coming from
this forum (which of course is most welcome) , as a matter
of fact i always relied on constructive discussion forums (just like
this one) to help me enhance my software as users are always able
to point me directions that i wasn't able to see in the first place.
Regarding the subject that the high precision pixelshader test result
is the same like the low precision , for ATI this is reasonable because
it has only one presision mode (24 bit) , so for high precision the
ATI card will execute 24 bits of precision , and for low precision it will
also execute 24 bits but will clamp the results afterwords to produce
16 bits of precision.
For Nvidia cards things are different , they support 12 bit , 16 bit , 32 bit
level of precision , so there must be a difference despite of if the
shader compiler used or not , because the shader compiler only change
the ordering of the execution of the program instruction but never changes
the precision , (or at least that's what nvidia says) , i wasn't able to test
the benchmark in GeforceFX cards because i only have a Radeon 9700 pro
,Radean 8500,Geforce4 TI4600 and Geforce3 TI500 at the moment ,
when i received reports that the GeforceFX is giving similer results i had
a closer look at the shader code and discoverd a bug in the high precision
pixelshader test led to it is always executing as low precision , i deeply
apologize for this mistake as the fact that i don't have GeforceFX and the
results should be identical on the ATI card of mine prevented me from
detecting this error before Release , now i corrected the shader code
along with some general modifications for the program based on the
very helpfull user opinions , and posted a new Release 1.1
(which you can download it at http://www.aureliosoft.com/download.html)
ofcourse the results could be the same after the modification as well
if any GeforceFX owner can post results here it will be of a great help
Also if anyone has comments or suggestins on the benchmark or the
program itself they will be wellcomed.
I want to thank you all for spending your valuable time for this discussion
it was realy of a great help , and i wish we could share more constructive thoughts ,
Ahmed Bahaa
Deathlike2
Regular
If what is said is true, then the results of R3XX tests should be the same, and there is a obvious difference in performance when tests are run with the FX.
Deathlike2
Regular
You probably aren't. I believe that the FX5800/5600/5200 (those that are based on the NV30, not the NV35) force FP16 at all times (someone could probably verify that). That's why you aren't getting different results.
Similar threads
