| VSA100 vs.
Voodoo3 Originally posted September 27, 2001 at x3dfx
"Discussion"
All performance graphs
courtesy of Bonehed.
Well, as promised I have spent
countless hours doing more benchmarks on obsolete hardware. All for the entertainment
purposes of our valued patrons here at x3dfx. The discussion arose as to whether or not a
standard Voodoo3 chip was as fast as a single VSA100. I knew from experience that the
Voodoo3 3000AGP in my secondary rig was much less powerful than my V5 5500, but I had
never really benchmarked the V5 in single chip mode much (whats the point when you've got
dual GPU's right?) :)
What I did was do all testing with my V5 5500 in my main gaming rig, then popped it out
and install my V3 3000 AGP using the same drivers. V5 drivers work fine for the V3, and I
felt like this would help eliminate differing driver versions as being causation for
differing results.
So here are my findings, get ready for some serious number crunching. You might be
surprised by some of my conclusions after doing these extensive tests, I know I was.
i440BX P3 @ 900MHz/112FSB
256MB Corsair PC100 @ CAS2
Quantum Fireball AS 7200RPM 20GB
V5 5500 AGP @ 166MHz
V3 3000 AGP @ 166MHz & 200MHz
Win98SE, IE5.01SP2, DX8.0a
Iceman 1.07.01 (combination of last official 3dfx release + x3dfx), default settings (post
filter enabled)
GLIDE: Unreal Tournament v4.36, Cityintro demo (timedemo 1), highest detail settings and
dynamic lighting, high quality sound.
OpenGL: Quake3 Arena v1.17, Demo001, default "normal" settings (16bit
color/textures), high quality sound.
D3D: 3DMark2000, default benchmark (1024x768x16bit)
Unreal Tournament
V5 5500 (fastest performance)
640 x 480 = 74.75 (Highest=131.89, Lowest=48.69)
1024 x 768 = 71.83 (Highest=120.81, Lowest=46.36)
1600 x 1200 = 52.33 (Highest=102.89, Lowest=31.12)
Wow, Unreal Tournament completely playable at
1600x1200 in GLIDE, never dipping below 30FPS.
V5 5500 (single chip only)
640 x 480 = 73.80 (Highest=125.82, Lowest=47.08)
1024 x 768 = 57.35 (Highest=112.68, Lowest=31.19)
1600 x 1200 = 24.86 (Highest=47.23, Lowest=13.85)
Ouch, the loss of one of the VSA100's really
brings down the FPS in higher resolutions. Notice how the frame rates double at fill rate
limitations, like at 1600x1200 resolution. Similar performance can be expected in FSAA.
V3 3000 (166MHz)
640 x 480 = 76.47 (Highest=127.96, Lowest=49.95)
1024 x 768 = 57.17 (Highest=108.80, Lowest=30.60)
1600 x 1200 = 27.51 (Highest=52.52, Lowest=15.39)
V3 3000 (200MHz)
1024 x 768 = 65.73
1600 x 1200 = 31.39
What a surprise! The V3 at the
exact same clock speed (166Mhz) is equal to or faster in this GLIDE game than a single
VSA100, and at 200MHz the V3 beats it hands down. I believe the V3 is probably more
efficient rendering in this API than the VSA100 for whatever reason, which explains why it
is slightly faster in all resolutions other than 1024x768 (in which it is equal for all
practical purposes). This should translate into the the V3 3000 AGP being pretty much
equal to a V4 4500 in GLIDE games.
Quake3
V5 5500 (fastest performance)
640 x 480 = 100.4
1024 x 768 = 94.2
1600 x 1200 = 44.3
Pretty impressive, yes? Nearly 100FPS at
1024x768 resolution in 22bit equivalent color dithering, and this on a sub-1GHz system.
Who says the V5 can't do OpenGL?
V5 5500 (single chip only)
640 x 480 = 97.7
1024 x 768 = 57.7 WGL = 57.8
1600 x 1200 = 22.7 WGL = 23.0
Once again, we see performance cut
in half with the loss of one VSA100 chip. WickedGL "hi-resolution" driver had
nearly identical performance to the 3dfx driver, which just goes to show 3dfx finally did
it right with the V5 5500.
V3 3000 (166MHz)
640 x 480 = 92.1
1024 x 768 = 42.4
1600 x 1200 = 18.3
V3 3000 (200MHz)
1024 x 768 = 52.8 WGL = 57.9
1600 x 1200 = 21.9 WGL = 24.4
Surprised again? I was! It is clear
that the V3 suffers in OpenGL performance compared to the VSA100, but overclocking up to
200MHz and using the WickedGL "high resolution" driver once again puts us at or
above single VSA100 performance levels. Amazing.
3DMark2000 (tip: with a P3
you may get a higher score turning off "geometry assist.")
V5 5500 (fastest performance) = 4495
Game 1: L = 91.1, M = 57.8, H = 24.4
Game 2: L = 104.4, M = 61.8, H = 35.2
Fill Rate (single): 516.4 MTexels/s
Fill Rate (multi): 590.9 MTexels/s
Texture Rendering: 8MB = 324.9, 16MB = 316.2, 32MB = 239.0, 64MB = 130.3
V5 5500 (single chip only) = 2967
Game 1: L = 56.1, M = 39.2, H = 17.9
Game 2: L = 54.0, M = 47.1, H = 34.0
Fill Rate (single): 253.7 MTexels/s
Fill Rate (multi): 310.5 MTexels/s
Texture Rendering: 8MB = 149.8, 16MB = 144.6, 32MB = 120.5, 64MB = 111.9
V3 3000 (166MHz) = 3193
Game 1: L = 61.3, M = 44.0, H = 19.9
Game 2: L = 57.1, M = 49.3, H = 34.6
Fill Rate (single): 159.2 MTexels/s
Fill Rate (multi): 312.6 MTexels/s
Texture Rendering: 8MB = 136.5, 16MB = 134.0, 32MB = 116.0, 64MB = 29.5
The V3 3000 gets a higher 3dMark score than
the single VSA100 at the same clock speed, on the exact same system. Notice the
texture rendering speed is faster on the single VSA100, however, and this will probably
mean faster performance in more texture intensive games and applications.
Conclusions: The Voodoo3 performs
just as well if not better than the VSA100 clock for clock in GLIDE and D3D. OpenGL is a
different story, and the VSA100 is superior here, although the V3 performs similarly when
clocked up to 200MHz and using the WickedGL. The VSA100 does have some advantages, many of
which are readily apparent. The additions of FSAA, large texture support, and 32bit color
in games may provide more longevity for the owners of V4 4500 cards, but in speed the
single VSA100 is fairly equal to the older V3. Thank you for watching, good night. |
