Luminescent
Veteran
I was just wondering how the R200's pipelines were enhanced to fit in the R9000 product line-up. Does anyone know how they're more efficient?
How were the RV250's pipelines improved over the R200's?
- Thread starter Luminescent
- Start date
Nebuchadnezzar
Legend
What the hell you talking?? There weren't no enhachement, R9000 is a cut-down R8500 ...
Luminescent
Veteran
Trust me, there were some types of efficiency enhancements to the vertex and pixel program processors (maybe some design philosophies borrowed from the R300?). The RV250 has been somewhat cut down in comparison to the R200: 1 tmu per pipeline, 1 vertex program processor, and no fixed-function T&L pipeline.
The changes to the 8500 core (R200) which resulted in the 9000 (RV250) were that the RV250 had 1 vertex shader unit removed (leaving one), which results in slower vertex processing power. This was semi-offset (not really) by the 9000 having an enhanced triangle setup engine over the 8500.
The 2 TMUs/pipe from the 8500 were replaced with a more advanced single TMU/pipe which came from the 9700. The single TMU can do 6 textures/pass instead of needing the 2 TMUs/pipe the 8500 has to have to achieve 6 textures per pass.
The 9000 also got more powerful pixel shaders from the 9700 instead of the pixel shader in the 8500. (can do things like the video processing the 9700 PS can do, which the 8500 can't, as well as being speedier than the 8500 PS)
Also the 9000 got the memory bus from the 9700 not the 8500.
All of that lowered the transistor count significantly, and hence the 9000 is a much cheaper chip than the 8500 to produce, while trading for some performance loss in some cases and maintaining performance in other cases (PS performance in particular here).
Edit: Oh yea, the 9000 aside from the pipeline changes also had a 2nd ramdac integrated into the chip, removing the need for an external ramdac for support for 2 vga monitors, which caused some confusion with some 3rd party 8500 boards. (and also helping reduce board costs for the 9000)
The 2 TMUs/pipe from the 8500 were replaced with a more advanced single TMU/pipe which came from the 9700. The single TMU can do 6 textures/pass instead of needing the 2 TMUs/pipe the 8500 has to have to achieve 6 textures per pass.
The 9000 also got more powerful pixel shaders from the 9700 instead of the pixel shader in the 8500. (can do things like the video processing the 9700 PS can do, which the 8500 can't, as well as being speedier than the 8500 PS)
Also the 9000 got the memory bus from the 9700 not the 8500.
All of that lowered the transistor count significantly, and hence the 9000 is a much cheaper chip than the 8500 to produce, while trading for some performance loss in some cases and maintaining performance in other cases (PS performance in particular here).
Edit: Oh yea, the 9000 aside from the pipeline changes also had a 2nd ramdac integrated into the chip, removing the need for an external ramdac for support for 2 vga monitors, which caused some confusion with some 3rd party 8500 boards. (and also helping reduce board costs for the 9000)
Luminescent
Veteran
Any word on what happened to the vertex shader? Supposedly it also borrowed some tech from the R300. According to Tom's hardware:
"The second change has to do with the vertex shader. It wasn't till now that ATI has admitted that the Radeon 8500 contains two vertex shader units, similar to the GeForce4 Ti. The reason for keeping quiet on this was more because of marketing concerns rather than understatement. The R8500 had two shader units more than NVIDIA's GeForce3. With the R9000, these units were re-worked and optimized. If the Canadian PR department is to be believed, then the new vertex shaders contain many optimizations that are also found in the R300 design."
Any speculations on what could have been changed here in the pipeline? How were the R200's vertex pipes limited? How do guys think they were improved in this design?
"The second change has to do with the vertex shader. It wasn't till now that ATI has admitted that the Radeon 8500 contains two vertex shader units, similar to the GeForce4 Ti. The reason for keeping quiet on this was more because of marketing concerns rather than understatement. The R8500 had two shader units more than NVIDIA's GeForce3. With the R9000, these units were re-worked and optimized. If the Canadian PR department is to be believed, then the new vertex shaders contain many optimizations that are also found in the R300 design."
Any speculations on what could have been changed here in the pipeline? How were the R200's vertex pipes limited? How do guys think they were improved in this design?
Luminescent said:
I don't know of any specific changes to the VS in the 9000 other than that it only has one, and that the 9000 got an "enhanced triangle setup engine" to go with it.
That doesn't mean there isn't something else going on under the hood that made the single 9000 unit perhaps more efficient or capable over the individual units from the 8500(9100)... but over all VS performance is better on the dual 8500(9100) VS than on the 9000 VS.
Luminescent
Veteran
If you read the result of this Max Payne benchmark: http://www.beyond3d.com/reviews/ati/radeon9000pro/index.php?page=page3.inc
from this here site, you will see that for being only 1 vertex shader, it is able to keep up rather nicely with the 2 VS's from the R200.
from this here site, you will see that for being only 1 vertex shader, it is able to keep up rather nicely with the 2 VS's from the R200.
Luminescent
Veteran
So I guess that double the texture cache (4k vs. 2k) explains why the RV250 can double the amount of texture passes per pixel pipeline (6 vs. 3) in comparison to the R200?
I am also guessing, if MuFu's list represents the only enhancements to the core, that the R200 only used one vertex pipeline (assuming no hardwired T&L unit) for hardwired T&L emulation, which explains the similarities in performance visible in the Max Payne benchmarks (linked somewhere above) between the R200 and the RV250 (both were only using 1 vertex engine).
I am also guessing, if MuFu's list represents the only enhancements to the core, that the R200 only used one vertex pipeline (assuming no hardwired T&L unit) for hardwired T&L emulation, which explains the similarities in performance visible in the Max Payne benchmarks (linked somewhere above) between the R200 and the RV250 (both were only using 1 vertex engine).
radar1200gs
Regular
The simple answer is that the RV250 is not more efficient than R200.
The entire 8500/9000 situation is about to take a hilarious and farcical turn anyway:
http://www.tech-report.com/onearticle.x/4455
Can it be that they r introducing R9000 Pro with 2 vertex shaders and calling it 9100?
I mean it would make it comparable to 8500 right? but it should perform better than 8500 owing to more optimized hardware and at less number of transistors. I am stating that cause the 2nd vertex shader in ti series did not add more than 10 million transistors if this is true it should still be cheaper to produce than 8500 (60 million transistors) and perform better than 8500
.
Anyway I heard that 9100 was actually RV280 the 8x agp refresh part of RV250 which seems more likely to me than renaming 8500 to 9100. (This is my basis for the speculation that it might also add a 2nd VS unit)
Just my 2 cents.
I mean it would make it comparable to 8500 right? but it should perform better than 8500 owing to more optimized hardware and at less number of transistors. I am stating that cause the 2nd vertex shader in ti series did not add more than 10 million transistors if this is true it should still be cheaper to produce than 8500 (60 million transistors) and perform better than 8500
.Anyway I heard that 9100 was actually RV280 the 8x agp refresh part of RV250 which seems more likely to me than renaming 8500 to 9100. (This is my basis for the speculation that it might also add a 2nd VS unit)
Just my 2 cents.
I doubt ATi would go through all the trouble of creating a new RV250 spin with only an extra VS. Likely they're just trying to clear out 8500 stock at maximal prices.
I'm looking for a budget card--who knows the best deal on a 64MB 8500/LE? Best I could find was $72 for a 250/230 card at NewEgg. I'm not sure if it's worth it, considering I can get a Tyan 275/275 9000 Pro for $87 @ TheNerds.net, which includes a DVI adapter and WinDVD.
I'm looking for a budget card--who knows the best deal on a 64MB 8500/LE? Best I could find was $72 for a 250/230 card at NewEgg. I'm not sure if it's worth it, considering I can get a Tyan 275/275 9000 Pro for $87 @ TheNerds.net, which includes a DVI adapter and WinDVD.
It depends on what you're going to benchmark. I can set up situations were either card wins hands down, but generally, a 8k5 LE is a bit more speedy than even a R9000pro @275MHz.
[edit]
At the same core-freq, pipelines of the R9000 have been improved, esp. short and easy shaders (including PS-EMBM), i.e. the most likely first-to-be-used in actual games, perform better on an R9000 (quite likely due to an increased/optimized texture cache), while in situations with longer shaders the R8k5 can keep it's edge.
[edit]
Code:
R9000 R8500LE
Environment Bump Mapping 127.8 fps 106.5 fps
DOT3 Bump Mapping 72.8 fps 81.3 fps
Vertex Shader 58.3 fps 80.4 fps
Pixel Shader 102.0 fps 93.0 fps
Advanced Pixel Shader 69.2 fps 73.5 fps
Point Sprites 16.3 MSprites/s 25.9 MSprites/sAt the same core-freq, pipelines of the R9000 have been improved, esp. short and easy shaders (including PS-EMBM), i.e. the most likely first-to-be-used in actual games, perform better on an R9000 (quite likely due to an increased/optimized texture cache), while in situations with longer shaders the R8k5 can keep it's edge.
