Voodoo Shack


voodoo6000_table.jpg (53436 bytes)



Image Archive


x3dfx Discussion




Voodoo Shack: When Voodoo is what you do™

3dfxtears_small.gif (2175 bytes)

Voodoo5 CPU Scalability  

benchmarks originally posted November 20, 2001 at x3dfx "Discussion" by Teg "Solidstate"

There has been a great deal of debate on the x3dfx forums as of late about the Voodoo5 and it's ability to "scale" up with more powerful CPU's.  There is no doubt that the Voodoo5 has great scaling capabilities, as many people found when upgrading to more powerful T-Birds in the 1.XXGHz.   But how long will the party last?  How far can the V5 5500 actually continue to improve with a processor upgrade?

This was exactly the question in mind as Teg (Solidstate on the x3dfx boards) sat down to do 150 or so benchmarks with his V5 to find the answer.  Teg basically took his AMD T-Bird, unlocked the multiplier, and   set it at CPU speeds ranging from 600MHz - 1.4GHz at both 100 and 133FSB, with and without FSAA.  He chose to benchmark running Q3A's Four.66_DM demo.  Here's the system specifics:

AMD Thunderbird AYHJA Y
Abit KG7 Raid
256mb Crucial pc2100 DDR
Voodoo 5 5500 AGP (1.18bios)
SolidState Win2k driver 1.01.00beta
Windows 2000 SP2 IE6
Settings - Tools
AGP Command Fifo - Auto Detect
Refresh Optimization - Enable
VIA Chipset - Performance
3D Filter Quality - High
Alpha Blending - Sharper
Depth Precision (16/32) - Faster/Faster
Anti-Aliasing - Varies
Guardband Clipping - Enabled
Hidden Surface Removal - Disabled
Legacy Texture Compression - Enabled
Level of Detail Bias - -0.50
Maximum Buffered Frames - 1
MIP Map Dithering - Disabled
Rendering Color Depth - Software Controlled
Triple Buffer - Disabled
Vertical Sync - Disabled
Settings - Quake 3
s_initsound 0
com_maxfps 300
GL Extensions - On
Video Mode - Varies
Color Depth - Varies
Fullscreen - On
Lighting - Lightmap
Geometric Detail - Medium
Texture Detail - 66%
Texture Quality - 16bit
Texture Filter - Trilinear

And Now For The Results (click for larger image):


100fsbCPU.gif (10322 bytes)



133fsbCPU.gif (11008 bytes)



100fsbCPUFSAA.gif (10951 bytes)



133fsbCPUFSAA.gif (10653 bytes)


So, what does all this mean?   Well, the results of these benchmarks make it pretty clear that the V5 5500 has excellent ability to scale with faster CPU's at lower resolutions.  The V5 5500 may actually continue to get faster in resolutions of 1024x768x16bit and lower with a faster CPU, possibly even up to and beyond 2GHz.  But at 1024x768x32bit graphics, it has reached fill-rate limitations, and isn't going to get much faster no matter how fast the CPU is.  The same can also be said when using the card in any 2x FSAA setting over 640x480 resolution.  The card is severely fill-rate limited when using 4x FSAA at any resolution.  This all makes perfect sense, since the V5 simply doesn't have the memory bandwidth or the necessary fill-rate at these settings to continue to increase the speed, even with a very fast CPU.  The only way to increase frame-rate when the card has become fill-rate limited is to overclock it, since doing so increases both the available memory bandwidth and fill-rate.

Basically what is happening is that at these settings, the card is using all it's resources to process the information, and the CPU is waiting for the graphics accelerator.  At lower resolutions where the card is at it's fastest, the SLI V5 is still fast enough to process the data and is waiting for the CPU, which is why at 640x480 the scores continue to increase with the faster CPU.  One day the V5 will meet it's match, and the CPU will be fast enough that no matter what the resolution is, the CPU will be waiting for the card.  This will be the day the V5 becomes completely fill-rate limited, and upgrading processors will have no effect whatsoever on performance increases.  

The great thing about all of this is that if you are like me, and run everything at 1024x768x16bit, you can still expect a nice little increase in performance when upgrading your CPU.  16bit graphics on the V5 are fantastic, and with the 3dfx post filter that dithers the 16bit colors into approximately 22bit (4million colors), you get good frame rates and great looking graphics.  The flip side to this is if you run your sims in 2x or 4x FSAA, as long as you have a 600MHz or faster CPU, there's really no need to upgrade your processor since you are getting maximum performance from the card as is.  I actually think both of these things are good depending on what type of user you are and system you have, and what type of games you play.

Special thanks goes out to Teg for taking the time to run these benchmarks and allowing me to post these results here.

Copyrightę 2002 Nightstormer Productions