NVIDIA GT200 Rumours & Speculation Thread

[Tchock]

Again, I doubt all your assumptions you have as to how the RV770 will perform unless ATI has somehow fixed how poorly they perform now. As it stands now, 64SP(320US) can't compare to 128SP by Nvidia. 2x(X2) ATI 64SP(640US) can just match that same 15m old 128SP by Nvidia. And you are some how coming to the conclusion that 96SP(480US and most likely the correct rumor) or your hopes 160SP(800US) is some how going to destroy the rumored so specs of either 256SP or 384SPs. either of which would work at worst on par to your high hopes situation.

Once you address the what-we-deem unbalanced ALU:ROP/TMU power while keeping the 64/320MADDs intact, "RV670" shouldn't be a bad deal. WAG of such chip would put it on par with G92/G80, and that's not even considering what RV770 is adding to the chip natively.

And comparing dual GPUs is just downright... out of line. Why don't you drag in the G92x2, 256SPs and a cumulative die area of >550mm^2 too if you want to make wrongs right?


Back OT, I suspect G92 @ 55 could have some stuff removed and compensated by higher clocks. Right now the config's rather rigid, but if 55 really scales well for them and gives them adequate headroom, 96/48 might not be a bad proposition for the "98GT"/whateverissimillarishSKU after all.

Power usage should go up with borderline clocks, but most consumers have clearly defined that as near-trivial.

 

[Ailuros]

And supposedly they even have something else on top of GT200 too....

Well that explains some weird sounding rumours in some recent newsblurbs, mentioning a G92b, a GT200 and a G100 all between today and Q4 this year.

It didn't take much for most of us to figure out that a 55nm/G92 would serve to cover temporarily the performance/mainstream segment of the market; the rest doesn't make much sense. I'll have a pretty damn hard time believing that any IHV would be able to release nowadays within the next 8-9 months 3 refreshes (or 2 refreshes and 1 new generation, whatever floats anyone's boat...). The latter two if those really exist under those codenames should logically belong to the same product family.

 

[trinibwoy]

This sounds quite similar to G80 at its early days when only 8800GTX/GTS were availiable.

The difference being that there was no brand spanking new G7x derivative to fill in the lower levels. G92 at 55nm is going to be a fearsome chip on its own so that bodes well for Nvidia's future mainstream offerings. Although I guess we're already there today with $200 8800GTS-512 deals.

 

[XMAN26]

Once you address the what-we-deem unbalanced ALU:ROP/TMU power while keeping the 64/320MADDs intact, "RV670" shouldn't be a bad deal. WAG of such chip would put it on par with G92/G80, and that's not even considering what RV770 is adding to the chip natively.

And comparing dual GPUs is just downright... out of line. Why don't you drag in the G92x2, 256SPs and a cumulative die area of >550mm^2 too if you want to make wrongs right?


Back OT, I suspect G92 @ 55 could have some stuff removed and compensated by higher clocks. Right now the config's rather rigid, but if 55 really scales well for them and gives them adequate headroom, 96/48 might not be a bad proposition for the "98GT"/whateverissimillarishSKU after all.

Power usage should go up with borderline clocks, but most consumers have clearly defined that as near-trivial.

I dont agree with you as a G92 GTS with 128SP and 16 ROPS can beat a single RV670. So again I dont see how he or you now can say the RV770 is going to clean house unless ATI addresses how their shaders work amoung other things.

 

[Megadrive1988]

I believe GT200 was first mentioned by FUDO, so I call bullshit. G100 is the most likely code name IMO.

That makes sense. GT200 would mean yet *another* change in the way Nvidia names / codenames its GPUs.




So it looks like as of now the general plan would be something like

near term: 9800 GX2 (of course still dual G80/G92 based) highend card for H1 2008

midterm: G100 / NV55 / GeForce 10 series / highend refresh GPU for mid/late 2008.
Then a year or more of G100 / NV55 respins, tweaks, shrinks, speedbumps etc through all most of 2009.

long term: true next-gen Nvidia architecture: NV60 / "G111" / GeForce 11 with DX11 / D3D11 / Shader Model 5.0 support for late 2009 or 2010, in time to compete with Intel's Larrabee, and will offer a window, a glimpse on what next-gen console graphics will be *like*, whether you compare an Nvidia based PS4 or an AMD/ATI based Xbox3.

 

[Magnum_Force]

hmmm

As regards to the sub topic of Nvidia Shaders versus ATi Shaders, lets not forget that Nvidia Shaders have there own clock domain, thus run MUCH faster than core clock.

RV670 (HD3780) - 775mhz (core/shader clock) x 320 (No. of shaders).

G92 (8800GTS 512mb) - 650/1650mhz (core/Shader clock) x 128

775 x 320 = 248000

1650 x 128 = 211200

SO, as you can see, on pure numbers (see below) there really isn't much in it between the HD3870 and the 8800GTS 512mb. Also note that the HD3870 also uses it's shaders to produce AA, so that will also take a percentage of it's Shader power (how much I have no clue).

Also, from reading these boards (you people are really rather clever ! ) I assume the way Nvidia's shaders work is somewhat more effecient in terms of Shader units being used per clock, atleast in real work, sub optimal conditions.

*Note: I'm not sure what this final number represents, but I THINK the theory of gauging general performance this way is fine (someone clever correct me if I'm wrong, thanks! ).

Feel free to punch holes in my first post thus making me retreat to the rock I crawled from under, never to post again .

 

[ShaidarHaran]

While the clock domain and number of shader processors certainly bares relevance in relation to the overall ability to execute shader code, it is important to also factor in the GPU's ability to schedule shader instructions for those SPs to work on. NV's SPs seem to have a higher utilization rate than ATi's, due to several factors which I lack the ability to explain properly, so I'll let someone else do that if anyone cares to.

 

[nutball]

near term: 9800 GX2 (of course still dual G80/G92 based) highend card for H1 2008

midterm: G100 / NV55 / GeForce 10 series / highend refresh GPU for mid/late 2008.
Then a year or more of G100 / NV55 respins, tweaks, shrinks, speedbumps etc through all most of 2009.

long term: true next-gen Nvidia architecture: NV60 / "G111" / GeForce 11 with DX11 / D3D11 / Shader Model 5.0 support for late 2009 or 2010, in time to compete with Intel's Larrabee, and will offer a window, a glimpse on what next-gen console graphics will be *like*, whether you compare an Nvidia based PS4 or an AMD/ATI based Xbox3.

Might I suggest that for at least 33.3% of that list you're just making stuff up?!

 

[Domell]

To answer your question with another question, how on God's green earth do you know that GT200/100 (or whatever it's called) "FPU" performance is 2x times higher?

I have read it on one German site (the same which rumoured that NVIDIA will show GT200 on Computex).

Bingo.

Jen-Hsung even said at the last conference: "we'll be driving the costs of G92 down over the next couple of quarters" which means that G92 will be around for at least 6 more months and a shrink to 55nm will certainly help driving down the costs of G92 down. GT200 will be the replacement of GF9800GX2 in the highend. G9x stack will fill everything below it... And supposedly they even have something else on top of GT200 too....

I only hope it doesn`t mean that it will be a DualGPU solution like GF9800GX2 but based on G92B.

BTW There are some rumours that GT200 will have about 1B transistors.

 

[LordEC911]

BTW There are some rumours that GT200 will have about 1B transistors.

MORE THAN.
1.2-1.6b seems the more commonly rumored number.

 

[Ailuros]

I have read it on one German site (the same which rumoured that NVIDIA will show GT200 on Computex).

Which proves what exactly?

Here's a link from another german site from the first post in this thread:

http://www.hardware-aktuell.com/viewnews.php?article=707

If I look at that BS it could imply 2.3 TFLOPs if ALUs would be identical to those on G80; that's over 4x times of that what a G80 can reach on paper.

I'll make it easy for you and call both sources bullshit and at the same time insist that sterile FLOP numbers mean jacks**t if one isn't aware of far more important internal details of any architecture.

I only hope it doesn`t mean that it will be a DualGPU solution like GF9800GX2 but based on G92B.
BTW There are some rumours that GT200 will have about 1B transistors.

The two chips on the 9800GX2 are adding up already to 1.5 B transistors. You should ask yourself if that supposed "GT200" will have ~1B or >1B transistors. Alternatively make up your mind what you want to believe after all; because I find it extremely hard to believe that any IHV can pack nowadays twice the floating point power into merely ~250M transistors.

 

[Megadrive1988]

Oh I wasn't aware that Jen-Hsun mentioned GT200 by name.

Okay well I too expect NV55 / GT200 to be over 1 billion transistors.

I assume this is still the GPU that was first rumored/mentioned almost a year ago, the one that was to provide nearly a TFLOP of performance and therefore 2 to 3 times the FP performance of G80, the one that was mistakenly said to be G92.

Given the unofficial delay from Q4 2007 to sometime in H2 2008, I'd now expect somewhat more than 1 TFLOP on a single new highend GPU.

 

[Ailuros]

I assume this is still the GPU that was first rumored/mentioned almost a year ago, the one that was to provide nearly a TFLOP of performance and therefore 2 to 3 times the FP performance of G80, the one that was mistakenly said to be G92.

I don't think so.

Given the unofficial delay from Q4 2007 to sometime in H2 2008, I'd now expect somewhat more than 1 TFLOP on a single new highend GPU.

Well if things would be that easy, they could wait another 6 months and make that 4 TFLOPs this time around.

 

[trinibwoy]

SO, as you can see, on pure numbers (see below) there really isn't much in it between the HD3870 and the 8800GTS 512mb. Also note that the HD3870 also uses it's shaders to produce AA, so that will also take a percentage of it's Shader power (how much I have no clue).

Yep, you're spot on. When this rumble first started people were talking about R600's shader power advantage and then people were trying to figure out whether to count G80's missing MUL etc etc.

Now, it's a bit more straightforward. Even discounting the MUL on G92, it has about 85% of RV670's pure MADD rate. Couple that with G92's presumed higher efficiency and it's more of a toss up as to who has the real advantage when it comes to shader processing.

Then again, G92 is far from shader bound so a shader comparison might not really matter once the dust settles.

 

[Jawed]

I assume this is still the GPU that was first rumored/mentioned almost a year ago, the one that was to provide nearly a TFLOP of performance and therefore 2 to 3 times the FP performance of G80, the one that was mistakenly said to be G92.

There's plenty of supporting evidence, so it seems like quite a safe assumption to me.

Given the unofficial delay from Q4 2007 to sometime in H2 2008, I'd now expect somewhat more than 1 TFLOP on a single new highend GPU.

Hmm, whereas I'm thinking a ~1.3 billion transistor chip might just be more trouble than NVidia was expecting. With current rumours putting it at being delayed by 7-10 months, perhaps more.

Jawed

 

[Arun]

If I look at that BS [...]

BS? That was the first leak that correctly hinted at 512-bit 1GHz GDDR3. I'd give them a little bit more credit than that myself!

 

[AnarchX]

I doubt that an end Q2/early Q3-solution will just reach ~128GB/s with 16 memory-chips, although 140-160GB/s and 1280MB with only 10 chips should be possible in this time frame, with savings in die-size.

from Hynx

 

[Arun]

Well, if you want 1024MiB/chip anyway, GDDR3 would be cheaper than GDDR5 to achieve 128GB/s of bandwidth... NVIDIA's strategy for memory is sometimes quite strange though, I'll grant you that. Maybe if GT200 was delayed, it was originally slated to be ready before GDDR5?

 

[Freak'n Big Panda]

Yep, you're spot on. When this rumble first started people were talking about R600's shader power advantage and then people were trying to figure out whether to count G80's missing MUL etc etc.

Now, it's a bit more straightforward. Even discounting the MUL on G92, it has about 85% of RV670's pure MADD rate. Couple that with G92's presumed higher efficiency and it's more of a toss up as to who has the real advantage when it comes to shader processing.

Then again, G92 is far from shader bound so a shader comparison might not really matter once the dust settles.

Spot on in that regard but the numbers in the post don't make any sense. Much better to calculate FLOPS and compare that way. You need to keep in mind that these shaders are just groups of ALUs arranged in SMID fashion in the case of RX6XX. For example:

RV670/R600: 5 MADD units per processing unit, 16x per cluster. 4 clusters for RV670/R600 chips. MADD (multiply add) is 2 floating point OPs so that's 160 FLOPs per cluster x 4 for the whole chip.

160x4x775MHz = 496 GFLOPs

G80/G92: 128 Shader units, supposedly able to accept one MADD and MUL instruction per clock giving 633.6 GFLOPs... not counting this 'missing MUL' we end up with 422.5 GFLOPs

For shader power calculate the FLOPs and then compare that way. That's the best way to sum it up in a single number.

 

[trinibwoy]

For shader power calculate the FLOPs and then compare that way. That's the best way to sum it up in a single number.

Well it's another way, not necessarily the best way in this particular case. His approach still works if you just look at the MADD rate and discount G92's MUL. To my knowledge nobody has been able to develop evidence that it's possible to consistently dual-issue the MUL.