maha_x
08-17-01, 04:04 AM
Ive been looking trough load of kyro2 reviews and the seem to get very different results.
I think I looked over 10 reviews. Only two of them found out that enabling Trilinear filtering can lead up to 30% hit in performance. And only one of them found out also that using texture compression in Q3A seemed to remove the perf. hit. After going trough some forums it began to look like if the mem bandwidth is to be blaimed. The mem bandwidth theory has some drawbacks. The same hit in FPS is same even at 16bit rendering, alltought more mem bandwidth should be available. This leads me to think that the architechture itself needs more clock cycles to do the Trilinear. But why texture compression nullifies this? Mayby the bandwidth bottleneck is in the Kyro2 mem controller where the data is still rendered in 32bit. this could explain why the perf. hit occurs similarly on both, 16 and 32bit modes, but is helped with text.compression.
Another issue found is the DX8 bug which keeps Kyro2 from rendering into textures. It can do it, but DX8 bugs, so all attempts to do it resort to do it in software. One popular use for rendering into textures is shadow generation. This hinders Kyro2's scores in Giants for example. Im not sure, but 3DM2k1 could be using this for shadows also. This will be fixed in DX8.1.
3DM2k1
3DM2k1 is giving out generally poor results troughout the line. There are several reasons that hinder the score.
Possibly render to texture shadows. Game 2? DX8 bug forces them to be done in software
Lack of DXT3 texture compression (DXT1 only). Textures are run w/o compression. The 30% hit from trilinear filtering kicks in.
DOT3 bumpmaps. They seemed like poison to Kyro2. Game1 High detail uses this.
Exactly these (Game1 hi-detail and Game2) were pointed out as the Kyro2 weak points. Luckyly the DXT3 compression is only usefull for translucent textures and is not such a big flaw. If 3DM2k1 was able combine the use of compressed and uncompressed textures, it could help Kyro2's score plenty. Also, using EMBM bumpmaps (more popular anyway) would help and the DX8 bug will be fixed.
Also a visual anomality was detected in the hi-poly test, where some of the polygons were missing. This seems to me like a problem in the size of the displaylist. A driver issue. This anomality was recorded by only one reviewer.
anyways, Im recieving my Kyro2 today or monday. Ill be using it till there is a good price/perf DX8 part out.
P.S: Id be interested to see if anyone with a Kyro2 could do some benchin in 3DM2k1 set to Bilinear filtering?
I think I looked over 10 reviews. Only two of them found out that enabling Trilinear filtering can lead up to 30% hit in performance. And only one of them found out also that using texture compression in Q3A seemed to remove the perf. hit. After going trough some forums it began to look like if the mem bandwidth is to be blaimed. The mem bandwidth theory has some drawbacks. The same hit in FPS is same even at 16bit rendering, alltought more mem bandwidth should be available. This leads me to think that the architechture itself needs more clock cycles to do the Trilinear. But why texture compression nullifies this? Mayby the bandwidth bottleneck is in the Kyro2 mem controller where the data is still rendered in 32bit. this could explain why the perf. hit occurs similarly on both, 16 and 32bit modes, but is helped with text.compression.
Another issue found is the DX8 bug which keeps Kyro2 from rendering into textures. It can do it, but DX8 bugs, so all attempts to do it resort to do it in software. One popular use for rendering into textures is shadow generation. This hinders Kyro2's scores in Giants for example. Im not sure, but 3DM2k1 could be using this for shadows also. This will be fixed in DX8.1.
3DM2k1
3DM2k1 is giving out generally poor results troughout the line. There are several reasons that hinder the score.
Possibly render to texture shadows. Game 2? DX8 bug forces them to be done in software
Lack of DXT3 texture compression (DXT1 only). Textures are run w/o compression. The 30% hit from trilinear filtering kicks in.
DOT3 bumpmaps. They seemed like poison to Kyro2. Game1 High detail uses this.
Exactly these (Game1 hi-detail and Game2) were pointed out as the Kyro2 weak points. Luckyly the DXT3 compression is only usefull for translucent textures and is not such a big flaw. If 3DM2k1 was able combine the use of compressed and uncompressed textures, it could help Kyro2's score plenty. Also, using EMBM bumpmaps (more popular anyway) would help and the DX8 bug will be fixed.
Also a visual anomality was detected in the hi-poly test, where some of the polygons were missing. This seems to me like a problem in the size of the displaylist. A driver issue. This anomality was recorded by only one reviewer.
anyways, Im recieving my Kyro2 today or monday. Ill be using it till there is a good price/perf DX8 part out.
P.S: Id be interested to see if anyone with a Kyro2 could do some benchin in 3DM2k1 set to Bilinear filtering?