Nvidia GT300 core: Speculation

[Arun]

Ah yes, tis fun looking at old postings:
http://forum.beyond3d.com/showthread.php?p=1230016#post1230016

Haha, sure is Of course, I never once claimed that wasn't speculative. It was even more speculative than what I just said Based on insider info from two separate sources claiming there was a 192-bit GDDR3, 384-bit GDDR3, and a GDDR5 part. Everything else was pure speculation. And obviously that base info turned out to be wrong.

GT216 has 48 ALUs Still waiting to find out the rest of its specifications.

I'm pretty sure it's 16 TMUs and 8 ROPs ala G84/G96. BTW, since we're talking about GT216: I'm pretty sure it'll be the GPU core for Ion2.

Very nice table by the way! Out of curiosity, how did you decide what was TA and what was TF? Proximity to what looked like L1 texture cache?

 

[Jawed]

Very nice table by the way! Out of curiosity, how did you decide what was TA and what was TF? Proximity to what looked like L1 texture cache?

Yep, I used what I thought was the block with most memory for TF (32 little blocks, they're cute). Still, it's all guesses, particularly as mine don't line up with either of the NVidia annotated die shots I have. But then both those (which are essentially the same in naming ALU and TEX) have very sloppy borders that don't line up with obvious block borders - which is why I ignored them.

Another interesting feature of the GT200 die shot is that it appears you can see a little routing repeater island in the centre of all the major blocks (except the ALUs) - i.e. logic that boosts signal on it's way from A to B. Well, that's my guess, anyway. These islands vary in size.

Jawed

 

[iwod]

LOL. Like his 'crazy' claims about NV's packaging problems?

Sorry, Charlie actually gets things right. He does have a bias, but yours appears to be worse.

DK

He is not simply bias... but EXTREMELY bias.

And he doesn't get thing"s" right, over the course he had less right claims then a my fingers count.

And he is getting worst and worst...Even Fudzilla is ten times better then SA.

 

[fellix]

Another interesting feature of the GT200 die shot is that it appears you can see a little routing repeater island in the centre of all the major blocks (except the ALUs) - i.e. logic that boosts signal on it's way from A to B. Well, that's my guess, anyway. These islands vary in size.

 

[Jawed]

Yeah, those.

 

[Groo The Wanderer]

OMG, thank you so much...I had absolutely no idea!

What's an "ogre"?

Charlie - What did you ever end up doing with your Segways?

I have both of them still. Segways are flimsy yet shatteringly expensive toys that break a lot. Currently the gen2 is dead (again), and not worth the money to fix. The gen1 is up and I use it a lot to go visit clients when I don't want to walk/have to get there quickly. It is quicker for me to take the segway to most places in downtown Minneapolis than it is to drive when you take parking into account.

I modded the gen1 with 4 foot cattle horns. When I strap them on, the twinks on the local segway tours (a company does tours of the Missisippi river/mill ruins/whatnot on them locally) really know who's boss.

-Charlie

 

[digitalwanderer]

If you have a pict of the horned Segway I'd love to see it! Thanks for the update, I've always sort of been curious about those things but they always looked just like very expensive and breakable toys to me.

 

[ShaidarHaran]

My boss rides one to work sometimes. It's a really odd sight seeing someone going full speed on one of those things.

 

[Ailuros]

Well yeah.. that was a yoke, I basically summed up all the rumours people seem to believe in now. i.e. GT300 is smaller than gt200@40nm yet double the performance. (i.e. about the size of G92 but packing double GT200 power)

I thought the rumors want D12U to be slightly smaller than GT200b@55nm? Ok I'll be generous and say that someone said that it's smaller than GT200 and the receiver automatically assumed that GT200b was meant. Still between the so far rumored 495 and 530mm2 the difference is so huge that those 7% difference between wannabe speculators that it'll probably decide as the one and only factor the chips success or failure...

pick a number from 16 to 2400

Would I disappoint you if I'd tell you that I expected far more?

 

[neliz]

Would I disappoint you if I'd tell you that I expected far more?

You mean AMD is playing the same tricks as last year

 

[Ailuros]

You mean AMD is playing the same tricks as last year

Does it work for you if I expect Cypress to be roughly at up to twice the RV770 performance level? It's a good starting point for each IHV's new generation as that's the typical goal. The hard part is to find out how they've gotten there.

 

[Domell]

If performance of Rv8xx (near P10k Vantage) is real don`t you think that NVIDIA will get a serious trouble with their GT300?
IMO to be successful GT300 should be about 30-40% faster than Rv870 or if there will be "GTX260 216 Core" or "GTX275" alike card based on GT3xx with similar price and performance better about 10-15% (like GT8800GT vs Rv670).

 

[Ailuros]

If performance of Rv8xx (near P10k Vantage) is real don`t you think that NVIDIA will get a serious trouble with their GT300?

That's the supposed Juniper score which is a mainstream GPU last time I checked. Wouldn't you want to think about that question again?

IMO to be successful GT300 should be about 30-40% faster than Rv870 or if there will be "GTX260 216 Core" or "GTX275" alike card based on GT3xx with similar price and performance better about 10-15% (like GT8800GT vs Rv670).

It should yes. Whether it'll be or not is still in the stars.

 

[Domell]

Yes, i know 10k i Vantage is a most probably Rv830 (or Rv840) score but for Rv870 we should get about 15k (if Rv870 is about 2X Rv830 in specs). Then GT300 shoul be about 2X or even more faster in real world situations than GTX285.

 

[Vincent]

That's the supposed Juniper score which is a mainstream GPU last time I checked. Wouldn't you want to think about that question again?



It should yes. Whether it'll be or not is still in the stars.

AA+AF performance is the most important factor, bar none.

 

[Ailuros]

Yes, i know 10k i Vantage is a most probably Rv830 (or Rv840) score but for Rv870 we should get about 15k (if Rv870 is about 2X Rv830 in specs). Then GT300 shoul be about 2X or even more faster in real world situations than GTX285.

Have we made up our minds yet if Cypress is a pure single core or a MCM after all? A GTX285 yields a tad less than P13k and a GTX295 a tad less than P18k. I still don't see why any hypothetical D12U should be "in trouble" in either/or case, since that supposed Juniper score lies right in between what a 4850 and a 4870 yields.

We're talking about the performance preset here which runs in a rather mediocre resolution to start with. I'll be generous though and call it shader limited as a synthetic test and try to remember that according to so far rumors D12U could end up having roughly 2.5x the arithmetic power than a GTX285; and not that shouldn't mean that I expect as much performance increase, but merely that I don't see anything that shows anything abnormal.

Alas if AMD's mainstream X11 GPUs wouldn't be able to reach at least X10 performance GPUs performance; no better make that for any graphics IHV.

 

[Ailuros]

AA+AF performance is the most important factor, bar none.

Problem is that we have so far only one hypothetical performance Vantage score from Juniper; an extreme preset score would be far more interesting.

 

[pjbliverpool]

I'm willing to give GT3xx a bit of breathing room as well if its going up against an AFR solution from AMD. Either a modest price premium, a small performance disadvantage or even a small combination of both would be acceptable to me if it means no input lag and more consistent performance.

 

[KonKort]

Problem is that we have so far only one hypothetical performance Vantage score from Juniper [...]

... and Redwood and Cypress score from tweakers.

You can read the whole story here.

 

[Ailuros]

... and Redwood and Cypress score from tweakers.

You can read the whole story here.

I meant we don't have a hypothetical relevant score from the competition to compare with.