| 3DMark 99 (1024x768x16)
Due to its lower memory bandwidth,
the Permedia3 inevitably suffers from a lower
performance at a higher screen resolution. So,
it's not surprising to see the RIVA TNT2 trash the
Permedia3 by 48-61%. Even the standard RIVA
TNT2 outperformed the overclocked Permedia3 by 37%!
|
The rasterizer score
is calculated from the Texture Rendering Speed
tests and the Fill Rate tests and shows
theoretical 3D performance. In 16-bit colour, the
Permedia3 is woefully outclassed by the RIVA TNT2.
Even overclocked, it could only come in 5%
slower than the standard RIVA TNT2!
|
This benchmark tests the
single-texturing performance of the card. With the
use of only one texture unit, the Permedia3 cannot
hope to beat the RIVA TNT2 at the resolution of
1024x768x16. In fact, it can only hope to be half as fast as
the RIVA TNT2!
|
Game 2 tests the card's
dual-texturing capability instead. However, even
though both texture units are working now, the Permedia3
only managed to close the performance gap
to 25% at standard clockspeed and 27%
overclocked. It's evident that the Permedia3's
slower framebuffer is dragging it down.
|
This benchmark tests the
single-texturing capability of the card. Its fillrate should be about 12%
slower than the standard RIVA TNT2 at standard
clockspeeds but equal it when overclocked.
However, the results showed that even when
overclocked, the Permedia3 was slower than the
standard RIVA TNT2 by 44%!Again,
this is proof that only one texture unit was in
use. That's why it's running at about half the
RIVA TNT2's performance.
|
In this benchmark of the card's
multi-texturing capability, the Permedia3 finally
showed the strength of its two texture units.
Although its fillrate is only 88% of the RIVA TNT2 at
standard clockspeeds, the Permedia3 managed to
match the RIVA TNT2's performance.
The same occured after
overclocking. Though the Permedia3 theoretically
would only have 86% of the overclocked RIVA TNT2's
fillrate, it was still able to perform as well as
the overclocked RIVA TNT2. Amazing...
|
Although 2MB of textures are being
used, the each 16-bit texture is smaller than a
comparative 32-bit texture. So, the Virtual
Texture engine possibly couldn't do much. That's why
the Permedia3 performed so badly in this test - it
was 31% slower than the RIVA TNT2. Memory
speed is much more important in this test.
|
When the texture size increased to
4MB, the Permedia3's Virtual Texture engine began
to show some effect. The RIVA TNT2 and the
Permedia3's performance gap dropped a little to
23% even though the texture size was small enough
that memory performance would be a major
performance factor.
|
With 8MB of textures, the Permedia3
finally came close to the RIVA TNT2 in texture
rendering speed. Now, the Permedia3 performed as
well as the RIVA TNT2. As the texture
load increases, the Permedia3 seems to be less
affected by AGP bus saturation.
|
Increase the texture size further
to 16MB and the Permedia3 is now actually faster
than the RIVA TNT2. In fact, it has a 25-28% advantage over
the RIVA TNT2.
|
However, when it came to 32MB
textures, the Virtual Texture completely failed
the Permedia3. The Permedia3 was only about a
third as fast as the RIVA TNT2's. It would appear that such a
massive load of textures has completely
overwhelmed 3Dlab's much vaunted Virtual Texture
engine.
|
In this forced 3-pass bump mapping
test, the Permedia3 performed rather poorly,
considering it has dedicated bump mapping hardware
that directly supports the emboss bump mapping
that 3DMark 99 uses. Even with the dedicated
hardware support for bump mapping, the Permedia3
was slower than the RIVA TNT2 by 36-37%.
|
The same thing here with the 2-pass
bump mapping test, except that the RIVA TNT2
further improved on its lead by outperforming the
Permedia3 by 83-84%!
|
With single-pass bump mapping, the
performance gap between the RIVA TNT2 and the
Permedia3 dropped only slightly to 35-36%. Even with its
triple-blend engine, the Permedia3 still couldn't
beat the RIVA TNT2. :(
|
The Permedia3 apparently can
perform point sample filtering about 10% faster
than it can do bilinear filtering. The RIVA TNT2,
on the other hand, only gained an extra percent of
performance with the simpler point sample
filtering mode.
|
While the RIVA TNT2 obviously
received a slight penalty using the trilinear
filtering mode (~1.5%), the Permedia3 surprisingly gained
a small performance benefit from using trilinear
filtering instead of bilinear filtering! I guess
3Dlabs must have optimized the Permedia3 for the
higher quality trilinear filtering.
|
Again, there's no support for
anisotropic filtering... so it's not surprising
the Permedia3 got a big N/A!
|
|
|
|
