9600GT SLi review

[BFG10K]

Surely there will be games that stress shaders more than others. There are games that need more than 64 shaders. But CoD4 was just the game we happened to try out.

If you ran the tests properly I'm convinced you could even get a difference in CoD 4. Give me some time to make a GPU limited demo and I'll see what I can find out.

You can use RivaTuner to disable 64 shaders on your Ultra and run your same tests

Um, that's exactly what I did. Those figures were from a stock Ultra and also with 64 SPs disabled.

My 8800GTS in CoD4 running at 1680x1050 all setting high. 4xAA.

See, there's the problem, You're running a low resolution in a relative primitive game engine, plus who knows what CPU you?ve got or what map you were running (chances are the map was lightweight anyway, as MP maps tend to be).

So of course disabling 64 SPs isn't going to do much given you're essentially testing your CPU.

 

[Keysplayr]

Originally posted by: BFG10K

Surely there will be games that stress shaders more than others. There are games that need more than 64 shaders. But CoD4 was just the game we happened to try out.

If you ran the tests properly I'm convinced you could even get a difference in CoD 4. Give me some time to make a GPU limited demo and I'll see what I can find out.

Ok sounds great. Thanks.

You can use RivaTuner to disable 64 shaders on your Ultra and run your same tests

Um, that's exactly what I did. Those figures were from a stock Ultra and also with 64 SPs disabled.

Oh, I didn't realize you actually disabled them. I thought you just took 9600GT scores. My bad. Be certain NOT to disable ROPs. I ran into severe garbled screens and BSOD's.

My 8800GTS in CoD4 running at 1680x1050 all setting high. 4xAA.

See, there's the problem, You're running a low resolution in a relative primitive game engine, plus who knows what CPU you?ve got or what map you were running (chances are the map was lightweight anyway, as MP maps tend to be).

Well, I just wanted to see exactly how many shaders I had to disable before noticing a degradation in performance. I know resolution matters, but for me 1680 x 1050 mattered.

So of course disabling 64 SPs isn't going to do much given you're essentially testing your CPU.

So you think that maybe 19x12 requires more than 64 shaders worth? Most likely. I can try it on my system and let you know what I get.

 

[AzN]

Originally posted by: BFG10K

Those numbers look ridiculous to me. What resolution is that?

It depends on the game, but they vary between 1600x1200 to 1920x1440. Also all of them were tested with 4xAA and TrMS except for Crysis which used 2xAA.

I think BFG tested uber high resolution where 64 shader was bottlenecking a bit more.

Well yeah, I tested the settings I actually play games at.

Testing low resolutions will simply move the bottleneck to the CPU, especially if it's a primitive engine like CoD 4. The tests I have seen so far have largely been flawed.

Also setting the shader core to half will simply not work as there?s a limited range it can be out from the core and if it?s too far RivaTuner will reset it to default.

No one is denying higher resolutions need more shader. It's doing more shader effects.

The fact remains however 9600gt does well compared to 8800gt because majority of the games out there 64sp is enough to get close 8800gt. As more games use more shader yes 9600gt will become weaker. I think I already mentioned this earlier in the Tweaktown thread.

 

[BFG10K]

No one is denying higher resolutions need more shader. It's doing more shader effects.

Well then, I guess you agree with me since you can't possibly be advocating using low resolutions to tests GPUs.

The fact remains however 9600gt does well compared to 8800gt because majority of the games out there 64sp is enough to get close 8800gt.

Depending on the tests the 8800 GT can be up to 30% faster than the 9600 GT.

Given the 9600 GT is clocked higher, the GT doesn't quite have double the shaders and it also has new drivers, it seems about right.

 

[AzN]

:laugh:

 

[Tempered81]

Originally posted by: BFG10K

No one is denying higher resolutions need more shader. It's doing more shader effects.

Well then, I guess you agree with me since you can't possibly be advocating using low resolutions to tests GPUs.

The fact remains however 9600gt does well compared to 8800gt because majority of the games out there 64sp is enough to get close 8800gt.

Depending on the tests the 8800 GT can be up to 30% faster than the 9600 GT.

Given the 9600 GT is clocked higher, the GT doesn't quite have double the shaders and it also has new drivers, it seems about right.

its about 50% faster at 19x1200 in bioshock.

 

[AzN]

Originally posted by: jaredpace

Originally posted by: BFG10K

No one is denying higher resolutions need more shader. It's doing more shader effects.

Well then, I guess you agree with me since you can't possibly be advocating using low resolutions to tests GPUs.

The fact remains however 9600gt does well compared to 8800gt because majority of the games out there 64sp is enough to get close 8800gt.

Depending on the tests the 8800 GT can be up to 30% faster than the 9600 GT.

Given the 9600 GT is clocked higher, the GT doesn't quite have double the shaders and it also has new drivers, it seems about right.

its about 50% faster at 19x1200 in bioshock.

Not according to Firingsquad.

http://www.firingsquad.com/har...performance/page14.asp

The thing with 8800 series and the 9600gt is that TMU is linked to SP. If you disable SP it might disable the TMU as well. Considering 8800gt does 8 textures per clock for every 16SP and BFG's ultra only does 4 textures per clock it might have more substantial disadvantage for the ultra but not 8800gt when SP is disabled because even if 64SP disabled for 8800gt it would still have texturing fillrate of an ultra.

Of course this also has to be tested. 3dmark texture fillrate test would show this.

 

[Tempered81]

Originally posted by: Azn

Originally posted by: jaredpace

Originally posted by: BFG10K

No one is denying higher resolutions need more shader. It's doing more shader effects.

Well then, I guess you agree with me since you can't possibly be advocating using low resolutions to tests GPUs.

The fact remains however 9600gt does well compared to 8800gt because majority of the games out there 64sp is enough to get close 8800gt.

Depending on the tests the 8800 GT can be up to 30% faster than the 9600 GT.

Given the 9600 GT is clocked higher, the GT doesn't quite have double the shaders and it also has new drivers, it seems about right.

its about 50% faster at 19x1200 in bioshock.

Not according to Firingsquad.

http://www.firingsquad.com/har...performance/page14.asp

The thing with 8800 series and the 9600gt is that TMU is linked to SP. If you disable SP it might disable the TMU as well. Considering 8800gt does 8 textures per clock for every 16SP and BFG's ultra only does 4 textures per clock it might have more substantial disadvantage for the ultra but not 8800gt when SP is disabled because even if 64SP disabled for 8800gt it would still have texturing fillrate of an ultra.

Of course this also has to be tested. 3dmark texture fillrate test would show this.

yeah thats using AA & AF, check it out without AA at 1920x1200

http://www.overclockersclub.co...iews/xfx_9600_gt/8.htm


edit: interesting now that we've posted comparisons of 1920 noaa and 1920 4xAA, I can't believe the difference in the performance hit the 8800GT takes over the 9600GT when enabling high AA. WTF I hope they work out those magical filters for the g92 as well.

 

[AzN]

Originally posted by: jaredpace

Originally posted by: Azn

Originally posted by: jaredpace

Originally posted by: BFG10K

No one is denying higher resolutions need more shader. It's doing more shader effects.

Well then, I guess you agree with me since you can't possibly be advocating using low resolutions to tests GPUs.

The fact remains however 9600gt does well compared to 8800gt because majority of the games out there 64sp is enough to get close 8800gt.

Depending on the tests the 8800 GT can be up to 30% faster than the 9600 GT.

Given the 9600 GT is clocked higher, the GT doesn't quite have double the shaders and it also has new drivers, it seems about right.

its about 50% faster at 19x1200 in bioshock.

Not according to Firingsquad.

http://www.firingsquad.com/har...performance/page14.asp

The thing with 8800 series and the 9600gt is that TMU is linked to SP. If you disable SP it might disable the TMU as well. Considering 8800gt does 8 textures per clock for every 16SP and BFG's ultra only does 4 textures per clock it might have more substantial disadvantage for the ultra but not 8800gt when SP is disabled because even if 64SP disabled for 8800gt it would still have texturing fillrate of an ultra.

Of course this also has to be tested. 3dmark texture fillrate test would show this.

yeah thats using AA & AF, check it out without AA at 1920x1200

http://www.overclockersclub.co...iews/xfx_9600_gt/8.htm


edit: interesting now that we've posted comparisons of 1920 noaa and 1920 4xAA, I can't believe the difference in the performance hit the 8800GT takes over the 9600GT when enabling high AA. WTF I hope they work out those magical filters for the g92 as well.

Considering BFG used AA on his test overclockerclub is irrelevant.

I wouldn't trust overclockerclub either. They had 8600gts beating out 3850 couple months back.

http://www.nvnews.net/reviews/...00_gt_ssc/page_7.shtml
Found another bench with bioshock. At @1920x1200 with no AA and it's 3fps difference.

 

[Tempered81]

how did BFG test a 8800GT when he has a 8800ultra?

 

[AzN]

He didn't test a 8800gt. He tested it on his ultra.

Considering an ultra does 4 textures per clock for every SP his texture fillrate would be close to 10000 Mtexels/S.. That would hamper his performance if TMU was disabled since TMU is linked to SP on these cards.

8800gt in the other hand would still have 18000 Mtexels/s even if 48 shader was disabled.

 

[Tempered81]

your tech level has surpassed mine Azn. I hear what you're saying, but won't understand till i see some benchies! In the meantime have a look at this GPU-z side-by-side comparison screenshot of a 8800GTS512 an 8800GTX and a 9800GTX and see if you can spot anything intresting....

http://www.xtremesystems.org/f...4&posted=1#post2797934

 

[AzN]

I'll be back. I'm going to test theory on my G84.

 

[AzN]

I just tested this with 3dmark mutil-texture fillrate test on my 8600gts G84 that does 8 textures per clock for 16SP.

8600gts default = 7000 Mtexels/s

8600gts with 16SP disabled = 3500 Mtexels/s

Since my 8600gts is linked to 2 64 bit memory controller it seems that my theoretical fillrate is never achieved. Probably 1 64bit memory controller linked for every 16SP that works independently.

So I was right along. BFG texture fillrate is cut into half when he disabled his 1/2 of SP which hampers his performance not because of his SP.

8800gt would see minimal impact even if 48SP was disabled because it would still have a fillrate of a 8800gtx that wouldn't be bottlenecked by memory bandwidth.

 

[Tempered81]

in laymans terms, what does this say about 8800gtx Vs. 9800gtx and 8800gt vs 9600gt?

and did you check out that screenshot?

 

[Cookie Monster]

Azn has a point BFG. Although let me explain it better. (but i still dont get your (Azn) obsession with fillrates! )

Your 8800Ultra, or G80 has 8 "clusters". Each cluster (ignoring caches, registers and such) has 16 scalar ALUs (2 pairs of 8), and 4 texture address/8 texture filtering units. Note that the TMUs are decoupled so i wouldn't use the word linked as Azn did , but they are part of that cluster. Anyway, each cluster has the ability to address 4 pixels per clock but 8 INT8 bilerps (FP16 half speed) per clock of fetch and filtering. I guess this was something that raised quite a few eye brows since this meant G80 will give you two bilerps per pixel address for "free". Basically the Ultra has 32 beefed up TMUs.

With the 8800GT or G92 (and G94), each cluster can address 8 pixels per clock i.e having a 1:1 ratio of TA/TFs unlike the 1:2 found in the G80. This results in no more "free" trilinear or 2x bileaner AF due to the 1:2 setup, but beefing up the address hardware i.e returning to the 1:1 ratio. Technically the G92 has 56 TMUs.

Now, if you disabled 1 clusters, this would mean your simulation of the GT has 28 TMUs compared to 56TMUs found on the G92. Same goes for the 9600GT (by disabling 4 clusters). 16 vs 32 TMUs. I think its fact that the ALUs found on G80/G92/G84 are all the same, but i was hoping nVIDIA to go dual MADD instead of their MADD+MUL. There are other variables (like memory bus difference/bandwidth, triangle setup etc) involved but i think the most important factor is the difference in the texturing performance.

One obvious reason why the RV670 (a full fledged R600 @55nm) is on par with the 9600GT (half of G80 @ 65nm). Its quite shocking thinking about it.

Anyway back to topic!

 

[AzN]

Like I said cookie you explain it better than me. English isn't my best subject.

FILLRATE! FILLRATE! FILLRATE!

 

[Cheex]

Just a quick comparison:

OLD - NEW
8800GTX - 9800GTX

GPU: G80 - G92
Technology: 90nm - 65nm
Die Size: 484mm^2 - 330mm^2
ROPs: 24 - 16
Shaders: 128 - 128
Pixel Fillrate: 13.8 GP/s - 10.8 GP/s
Texture Fillrate: 38.7 GT/s - 37.8 GT/s
Bus Width: 384 Bit - 256 Bit
Memory Size: 768MB - 512MB
Bandwidth: 86.4 GB/s - 70 GB/s
GPU Clock: 576 - 675
Memory Clock: 900 - 1100
Shader Clock: 1350 - 1688

 

[AzN]

OLD - NEW
8800GTX - 9800GTX

GPU: G80 - G92
Technology: 90nm - 65nm
Die Size: 484mm^2 - 330mm^2
ROPs: 24 - 16
Shaders: 128 - 128
Pixel Fillrate: 13.8 GP/s - 10.8 GP/s
Texture Fillrate: 18.4 GT/s - 37.8 GT/s
Bus Width: 384 Bit - 256 Bit
Memory Size: 768MB - 512MB
Bandwidth: 86.4 GB/s - 70 GB/s
GPU Clock: 576 - 675
Memory Clock: 900 - 1100
Shader Clock: 1350 - 1688

 

[Tempered81]

Originally posted by: Azn
OLD - NEW
8800GTX - 9800GTX

GPU: G80 - G92
Technology: 90nm - 65nm
Die Size: 484mm^2 - 330mm^2
ROPs: 24 - 16
Shaders: 128 - 128
Pixel Fillrate: 13.8 GP/s - 10.8 GP/s
Texture Fillrate: 18.4 GT/s - 37.8 GT/s
Bus Width: 384 Bit - 256 Bit
Memory Size: 768MB - 512MB
Bandwidth: 86.4 GB/s - 70 GB/s
GPU Clock: 576 - 675
Memory Clock: 900 - 1100
Shader Clock: 1350 - 1688

WT where did you get 18.4 from, and does this mean that the 9800gtx will perform 2x faster (with its 37.8GT/s) ?

 

[AzN]

Because it is 18.4 GT/s... 575 mhz x 32 tmu = 18400

8800gtx is 24 rop 32 tmu card. G92 is 16 rop 64 tmu.

No it would be bottlenecked by memory bandwidth just like G92.

 

[Tempered81]

Originally posted by: Azn
Because it is 18.4 GT/s...

\

Yes i know you said it is 18.4 GT/s. Where did you get the number 18.4 from? Is it the result of some division arithmetic or something? I'm going to reread your earlier posts about texture fillrate....

 

[Cheex]

WTF Azn?????

How so???

 

[AzN]

Originally posted by: Cheex
WTF Azn?????

How so???

http://techreport.com/articles.x/11211/5

G80 does 4 textures per clock for every 16SP cluster. 8800gtx has 128sp. 4x8=32 tmu

G92 does 8 textures per clock for every 16SP cluster. G92 8x8= 64 tmu

 

[BFG10K]

8800gt in the other hand would still have 18000 Mtexels/s even if 48 shader was disabled.

But again that texture fillrate advantage is only relevant in games not using FP rendering.

I tested Bioshock again (which does use FP rendering) using the lowest shader clock I could go to (1224 MHz, default 1512 MHz).

I got 47 fps vs 53 FPS, or a 13% performance gain for an increase of 24% in shader clock.

And again the difference would likely be higher if I wasn't running 4xAA combined with TrAA.

So I was right along. BFG texture fillrate is cut into half when he disabled his 1/2 of SP which hampers his performance not because of his SP.

But again you need to demonstrate that the reduction of texturing hurts performance more than the reduction of shading ability.

You would also need to demonstrate how your multi-texturing results are relevant in today?s games.

So far you have failed to do either while I have provided ample evidence to demonstrate shaders becoming more and more widespread compared to texturing.

We also have commentary from ATi in the article I linked to that confirms what I'm saying.