D10U-30 from nVidia

[Cookie Monster]

1billion transistors has been confirmed. And something about GT-200 being the actual name of the card. Links are at vr-zone.

 

[ajaidevsingh]

angry hampster, is right but the 768 bit for the extra AA and AF processes...!!! One hell of a hot rod that would make!!!

 

[Rusin]

Originally posted by: angry hampster

Originally posted by: Lonyo


512-768-bit bus width isn't going to happen, that's for sure.
Increasing bus width isn't a cost effective way of making a faster card when compared to using faster memory, especially in the long run if you want to be able to reduce costs.

It's not cost effective, but it's extraordinarily effective. We've already seen that G92 is quite limited by its 256 bit bus. A 448 or 512 bit bus with nearly a gig of memory would be fantastic. 768 is ridiculous at this point though

256-bit bus is enough for next gen if they would use GDDR5 memory.

 

[imported_Shaq]

But what kind of heatsink is this thing going to require? Will the 9800GTX part be good enough for another 80 watts?

 

[AzN]

Originally posted by: Rusin

Originally posted by: angry hampster

Originally posted by: Lonyo


512-768-bit bus width isn't going to happen, that's for sure.
Increasing bus width isn't a cost effective way of making a faster card when compared to using faster memory, especially in the long run if you want to be able to reduce costs.

It's not cost effective, but it's extraordinarily effective. We've already seen that G92 is quite limited by its 256 bit bus. A 448 or 512 bit bus with nearly a gig of memory would be fantastic. 768 is ridiculous at this point though

256-bit bus is enough for next gen if they would use GDDR5 memory.

Maybe a little bump but not for a real next generation. Not to mention you are limited to 16 ROP's. Is that even enough to play Crysis @ very high detail? 8800gts 512 is the pinacle of 16 pipe card. Sure it's limited by bandwidth but how much further will it scale with GDDR5? Little faster than 8800gtx with AA? Is that even worth it?

 

[Rusin]

If I remember correctly ROPs aren't that important in modern games like they used to. With AA texture mapping performance is the key

These GDDR5 samples that they had half a year ago were clocked at 2.5GHz and had similar power consumption as GDDR4. That's pretty much when you think that fastest GDDR4 solution is clocked at 1.2GHz and fastest GDDR3 solution is clocked at 1.18GHz. So theoretically you could get 160GB/s max transfere rate with 256-bit memory interface which is 25% upwards from HD2900 XT GDDR4 and it's 512-bit memory interface.

 

[Rusin]

Originally posted by: Shaq
But what kind of heatsink is this thing going to require? Will the 9800GTX part be good enough for another 80 watts?

Propably it requires cooler that has similar performance as 9800 GX2's cooler. Weren't 9800 GX2's TDP around 250W..meaning that this new generation would have similar TDP value.

 

[AzN]

Originally posted by: Rusin
If I remember correctly ROPs aren't that important in modern games like they used to. With AA texture mapping performance is the key

Where do you remember it from? Any links testing this theory?

What I've noticed and heard from the past is that Pixel fillrate has a lot to do with AA performance and high resolution tests as well. Everything comes into play. You can't just point to one thing and say it isn't important. Something might be more bottlenecking than other parts of chip. All cards are like this one way or another.

Look at 8800gtx with 24 ROP and 32TMU beating out 16 ROP 8800gt 56TMU or 8800gts 64TMU with AA in higher resolutions. Sure memory bandwidth has a lot to do with that but pixel fillrate comes into play as well when it comes to higher resolutions and or with AA.

Look at HD3870 as well. It's raw performance gets very close to 8800gt when you consider it only has 16tmu.

It just makes sense to raise everything so it's more balanced. Not beef up the memory bandwidth. It will be like 2900xt vs 3870.

 

[Rusin]

HD3870 has 30% better ROP performance compared to 8800 GT
HD3870 has 30% better ALU performance compared to 8800 GT
HD3870 has 25% bigger memory bandwidht compared to 8800 GT
Still 8800 GT wins HD3870 by 15-20%

Only area where 8800 GT outshines HD3870 on specs list is texture mapping performance: 8800 GT has 2.7 times bigger texture mapping-performance. If rumoured R700 specs are correct then HD3870 X2 (core 1050MHz 2x32 TMU units) will have smaller texture mapping performance than 9800 GX2 (core 600MHz 2x64 TMU).
------

You also need memory bandwidht when you use AA and/or higher resolutions. This is area where 8800 GTX shines..it has just right memory bandwidht.
----

When you overclock G92 cards you'll notice that you don't get much additional performance from it. If you overclock core like 20% you'll get 5% performance increase..even when in this example ROP-performance should increase by 20%.

 

[Keysplayr]

Originally posted by: bunnyfubbles
For those making comments about the ridiculous power requirements...keep in mind technologies such as Hybrid SLI...I certainly wouldn't mind a single GPU video card consume 250watts if it had performance to justify it - with something like Hybrid SLI, that power guzzling video card could be shut off entirely when not put to full use.

Originally posted by: TC91
im not liking how those 8pin connectors are becoming more common on video cards

people used to say the same thing about having any sort of external connector or dual slot (or more) cooling

Hybrid power is pretty cool. With a motherboard and graphics card to support it, the discrete graphics cards completely shut off and then utilizes the onboard graphics when not gaming.
Also, the onboard will support PureVideo2 as well.
So, even if the 250W power requirement is true, if you use hybrid power, over time you'll likely use less power than a system that "cant" shut off the discrete graphics card(s).

 

[superbooga]

Originally posted by: Rusin
HD3870 has 30% better ALU performance compared to 8800 GT

Where are you getting this number? Marketing from both companies state ~500 GFLOPS.

 

[Rusin]

Originally posted by: superbooga

Originally posted by: Rusin
HD3870 has 30% better ALU performance compared to 8800 GT

Where are you getting this number? Marketing from both companies state ~500 GFLOPS.

8800 GT: 1500MHz * 112 SP
HD3870: 775MHz * 320 SP

Sorry that difference is over 45% sorry for that mistake

 

[superbooga]

Originally posted by: Rusin

8800 GT: 1500MHz * 112 SP
HD3870: 775MHz * 320 SP

Sorry that difference is over 45% sorry for that mistake

You can't calculate it that way. As an example, Nvidia SP could do 3 madds per clock while an ATI SP could do 2 madds per clock.

To get the gflops numbers to make sense, Nvidia SP does 50% more instructions per clock.

 

[imported_Shaq]

Originally posted by: Rusin

Originally posted by: Shaq
But what kind of heatsink is this thing going to require? Will the 9800GTX part be good enough for another 80 watts?

Propably it requires cooler that has similar performance as 9800 GX2's cooler. Weren't 9800 GX2's TDP around 250W..meaning that this new generation would have similar TDP value.

It's supposed to be only 197 watts. You have to turn up the fan speed to 80% to keep it under 80C--with the side of my case off too. That's why I was wondering. Another 40 watts above that plus OC'ing and I don't think it will be good enough. Maybe we are into triple slot video card territory.

 

[Cookie Monster]

You cant compare a given architectural performance just looking at its fillrates. Just because the HD3870 has a higher pixel fillrate for example, doesn't equal to it having 30% ROP performance against the 8800GT. It doesn't make sense really.

 

[AzN]

Originally posted by: Rusin


When you overclock G92 cards you'll notice that you don't get much additional performance from it. If you overclock core like 20% you'll get 5% performance increase..even when in this example ROP-performance should increase by 20%.

You are probably testing memory deprived situations. When I raised the core clocks and shader clocks on my card it scaled properly. It just depends on what settings you are testing and what game. If you are testing with AA you are testing more with memory bandwidth and not the power of the core.

I've done a pixel fillrate with 3dmark here showing what kind of improvement I got just overclocking the core. Sure memory bandwidth holds back some but not as much as you think it does especially when AA is not applied.

3Dmark Pixel Fillrate Test

Reason why 9600gt does so well against 8800gt when AA is applied is the same reason. It's pixel fillrate is higher and memory bandwidth is same. When you test bunch of games with AA you are testing memory bandwidth and pixel advantage. Take out the AA than 8800gt is 15% more faster vs AA.

What I'm saying is you can't just point and say 16rop is all you need or we don't need any more pixel performance. A balanced card is just better.

If rumoured R700 specs are correct then HD3870 X2 (core 1050MHz 2x32 TMU units) will have smaller texture mapping performance than 9800 GX2 (core 600MHz 2x64 TMU).

You mean 4000 series x2? It'll whoop a gx2 for sure with only 2x32TMU.

 

[AzN]

Originally posted by: Cookie Monster
You cant compare a given architectural performance just looking at its fillrates. Just because the HD3870 has a higher pixel fillrate for example, doesn't equal to it having 30% ROP performance against the 8800GT. It doesn't make sense really.

You can't compare but it does add up when you compare performance.

 

[BFG10K]

If I remember correctly ROPs aren't that important in modern games like they used to. With AA texture mapping performance is the key

That statement doesn't make sense as ROPs are directly responsible for doing the AA resolve.

Texturing performance doesn't have any impact on AA unless you're doing super-sampling of some kind.

 

[Cookie Monster]

Im thinking it has to do with better memory management or some new compression method/algorithm behind G94 vs G92 when it comes to performing AA. G94s seem to get less of a hit when doing AA compared to the G92 counterparts. Could be also because of driver issues.

ROPs are hardly the bottleneck (G92 vs G80). What it comes down to is two things. Poor memory management (ALL G92 based cards take a nose dive in performance when it runs out of memory compared to ATi cards which can only mean that nV has a rather inefficent way of handling its memory) and a rather inefficient AA (Even with the G80 Ultra you still see has large AA performance hits, meaning that AA isnt too efficent either).

GT200 (NV55) will probably be a more refined G80 (NV50) with a 512bit memory interface. Something like NV40 was to G70/NV47. By the time we get to 2009~10, we will see nVIDIA's true next gen (NV60) in time for larabee which the whole intel vs nVIDIA will take place. Quite looking forward this.

 

[Rusin]

Originally posted by: Azn

You mean 4000 series x2? It'll whoop a gx2 for sure with only 2x32TMU.

Most likely yes, but since 9800 GX2 will be EOL'd in few months it's safe to say that Nvidia won't use 9800 GX2 to fight with AMD's next gen high end. GX2 already wins HD3870 X2 by something like 40-55%..so Nvidia doesn't have to improve that much as AMD is doing to stay ahead

There have been speculation that D10U-30 would look a lot like 9800 GX2 on one GPU. This could give them huge performance increase over 9800 GX2..when you remove pretty much all bottlenecking issues + SLI-issues.

 

[AzN]

Far as I know there isn't any special algorithm behind g94 vs G92. They both have the same tweaks. G92 performance improvement with 174.xx drivers shows this. However Geforce 9 series may have better SLI bridges considering it scales better with SLI.

http://en.expreview.com/2008/0...e-scaling-test/?page=5

 

[AzN]

Originally posted by: Rusin

Originally posted by: Azn

You mean 4000 series x2? It'll whoop a gx2 for sure with only 2x32TMU.

Most likely yes, but since 9800 GX2 will be EOL'd in few months it's safe to say that Nvidia won't use 9800 GX2 to fight with AMD's next gen high end. GX2 already wins HD3870 X2 by something like 40-55%..so Nvidia doesn't have to improve that much as AMD is doing to stay ahead

There have been speculation that D10U-30 would look a lot like 9800 GX2 on one GPU. This could give them huge performance increase over 9800 GX2..when you remove pretty much all bottlenecking issues + SLI-issues.

Yes I already speculated this on the first page.

It'll be something like this.

32 ROP, either 48 or 96 TMU, 192SP, 512bit memory controller 1024mb


28 ROP, either 40 or 80 TMU, 160SP, 448bit memory controller 896mb

It'll scale much better than what SLI can do. A single vidcard is just better.

 

[Rusin]

Well what about those power consumption rumours..that this would consume as much as 9800 GX2 does? With those specs this new card would have less processing units than 9800 GX2 [256 SP, 128 TMU, 32 ROP..basically 1/3 more]. Perhaps much higher clock frequencies?

 

[superbooga]

Current games aren't SP-limited at all.

 

[Rusin]

But TMU numbers are important (Hard to find any other reason why 8800 GT wins HD3870 by 15-20% margin). Main issue was that powerconsumption question.