The GT300/Fermi Thread - Part 2!

[RaV[666]]

I didn't say anything about heat or power usage, it is your words. I said: "nVidia doesn't have to beat ATi with more than ~10-20% all over the line up to make a success out of Fermi"

The problem is that it looks like it will be 20% better AT MOST.So depending on the game/benchmark/settings it will be like 0-20%.
Now take into consideration 5870 2gb OC ,then youre looking at MAYBE 10% AT MOST.Thats ugly for a product that probably will have power consumption more comparable with a 5970 ... .
It seems charlie was right with the INTENDED clocks of 750 too, with that clocks this card would be just about right.
Any solid info on price of 470 ?Maybe it will be more sensible offering, nvidia should have loads of not fully functional fermi cores ;-)

[zerazax]

I didn't say anything about heat or power usage, it is your words. I said: "nVidia doesn't have to beat ATi with more than ~10-20% all over the line up to make a success out of Fermi". What would 6 mounts have to say about this, in your opinion? should be 200% better and why?

Success? Based on what? Pure performance?

If I made a GPU that netted me a $15 profit while drawing 400w and heats up my coffee everyday, but is 20-30% faster than the 5970, but was half a year late and is in extremely limited numbers, is that a success? Okay that was an extreme example but my point is: You can't look at performance numbers in a vacuum and call the card a success.

Not to mention that if new rumors are true and the 480 is indeed 480sps instead of a full 512, then that's a failure to meet your own set requirements.

Furthermore, if the power/tdp numbers are true, it's going up against 5970s (don't bother rehashing 2v1 gpus, we are talking price and power figures since those are objective categorization points)

"Just" 10-20% (overall, which remains to be seen) after 6 months is dissapointing. It's not a flop like R600 (where it was slower) or NV30, but consider that GT200 vs RV770 had similar performance deltas but GT200 was almost a month earlier.

[thatdude90210]

The latest PC perspective podcast (~48min mark), Ryan Shrout says "the manufacturability of these chips is pathetic," and the yield is about 10-20% at best.

[Sam_oslo]

Success? Based on what? Pure performance?

If I made a GPU that netted me a $15 profit while drawing 400w and heats up my coffee everyday, but is 20-30% faster than the 5970, but was half a year late and is in extremely limited numbers, is that a success? Okay that was an extreme example but my point is: You can't look at performance numbers in a vacuum and call the card a success.

Not to mention that if new rumors are true and the 480 is indeed 480sps instead of a full 512, then that's a failure to meet your own set requirements.

Furthermore, if the power/tdp numbers are true, it's going up against 5970s (don't bother rehashing 2v1 gpus, we are talking price and power figures since those are objective categorization points)

"Just" 10-20% (overall, which remains to be seen) after 6 months is dissapointing. It's not a flop like R600 (where it was slower) or NV30, but consider that GT200 vs RV770 had similar performance deltas but GT200 was almost a month earlier.

You are repeating many rumors that nobody knows if they are true or not, yet.

You say 20% is disappointing after 6 months, but 20% lead would be the best business scenario for nVidia and make many people (with sans for high-end GPU) happy too.
You are avoiding to tel if it should be 200% better or what to satisfy you? and what those 6 mounts got to do with performance crown of the best performing GPU av the time?

[Solus Corvus]

But we already know its performance so there's no need to wait. We know Fermi does Tessellation well. The guy in the video even says it's what sets them apart from the competition. This is exactly what Charlie said. He also said the GTX480 is only 0-5% faster than the 5870. He's been right for months now, so I'm going by his word.

Like I said, there's no reason to buy this card unless you're a die-hard Nvidia fan or use GPGPU apps.

I'm not some knee-jerk Charlie hater like some of the people around here. I realize that he sometimes gets thing right and sometimes wrong. But I'm not just going to take his word on it. Doing so would be foolish, imo. We don't know about any of the test details of the numbers Charlie gave, even if they are right.

nVidia doesn't have to beat ATi with more than ~10-20% all over the line up to make a success out of Fermi. Don't forget that the 5870 with that hefting price tag (comperad to 5850), is only ~10-20% better than 5850.

If GTX 480 is ~10-20% better than 5870 then nVidia has a winner, for sure, even after 6 mounts. Unless ATi can release a new/refresh card to match/beat it at the launch time.

If it's only 10-20% faster then Nvidia is in trouble. That's well within reach of the 1Ghz 5870 cards that are being released and will probably be available in volume before Fermi is even released.

It's not like last gen where 4890 was released much later and still couldn't match a 285. Instead you'd have a situation where the top single GPU NV card is matched in performance by something that came out before it did. That would be much worse then last gen for NV.

And that's just talking partner OCed 5870s that we already know are coming out. ATI could start binning or do a respin and introduce even faster 5890s.

The problem is that it looks like it will be 20% better AT MOST.So depending on the game/benchmark/settings it will be like 0-20%.

In Unigine there appear to be sequences where the 480 is ~60% faster then a 5870. That is a very significant lead. But Unigine Heaven is a highly contrived scenario with lots of tessellation and not much shaders. Performance in more tess/shader balanced scenarios remains to be seen, not to mention realistic AA/AF settings.

[Sam_oslo]

..

If it's only 10-20% faster then Nvidia is in trouble. That's well within reach of the 1Ghz 5870 cards that are being released and will probably be available in volume before Fermi is even released.

It's not like last gen where 4890 was released much later and still couldn't match a 285. Instead you'd have a situation where the top single GPU NV card is matched in performance by something that came out before it did. That would be much worse then last gen for NV.

And that's just talking partner OCed 5870s that we already know are coming out. ATI could start binning or do a respin and introduce even faster 5890s.


....

I have already said it in 2 posts back:

..
If GTX 480 is ~10-20% better than 5870 then nVidia has a winner, for sure, even after 6 mounts. Unless ATi can release a new/refresh card to match/beat it at the launch time.

[RaV[666]]

Taking into consideration that it has 40% more transistors than cypress, and also has wider (more expensive) memory bus.
How can you call that "Best business scenario" ?
I bet better business scenario for nvidia would be 40% better performance for 50% more cash and wide availability about 6 months ago.

edit:

In Unigine there appear to be sequences where the 480 is ~60% faster then a 5870. That is a very significant lead. But Unigine Heaven is a highly contrived scenario with lots of tessellation and not much shaders. Performance in more tess/shader balanced scenarios remains to be seen, not to mention realistic AA/AF settings.

In the same slide you see not only spikes downard but also being on par with fermi, so in this benchmark end result wont be 60%.Nitpicking best case benchmark and even best case scenario IN the said benchmark isnt really fair isnt it ?

[Sam_oslo]

mVidia and ATi both has been asking for your short for just 10-20% performance increase all the time. You can't afford a 40% increase for Fermi, unless it becomes a really bad business for nVidia. The same would be the case for ATi in it's turn too.

[Solus Corvus]

I have already said it in 2 posts back:

ATI doesn't have to do anything. ATI board partners have already announced 1GHz cards. We have even seen more of them then we have seen real Fermi cards so far.

If fermi is only 20% faster then a regular 5870 then it's going to be matched by cards we already know exist. There is no question of "IF ATI does X" because that isn't necessary, that would just be rubbing salt in the wounds.

In the same slide you see not only spikes downard but also being on par with fermi, so in this benchmark end result wont be 60%.Nitpicking best case benchmark and even best case scenario IN the said benchmark isnt really fair isnt it ?

No, it's not really fair. I'd like to point out that in the non-dragon sections of Unigine a stock 5870 seems to match a 480 - so Charlie may in fact be correct that a 480 is roughly equal to a 5870 in most games.

I just pointed out the 60% figure to show that you aren't technically correct about 20% being the max and that there will always corner cases for the fanboys to haggle over.

[Sam_oslo]

Lets hope you are right. That would be the best scenario for us frequent consumers av GPU. If those 1GHZ 5870 gets ready and can match/beat 480 at launch, then we are looking at the era of $499 for the top tog.

Good and cheap GPU from both ATi and GPU would be flying everywhere then.

[Solus Corvus]

If Nvidia's top card only matches a 5870 then I don't see many price reductions for 5970 coming.

[zerazax]

You are repeating many rumors that nobody knows if they are true or not, yet.

You say 20% is disappointing after 6 months, but 20% lead would be the best business scenario for nVidia and make many people (with sans for high-end GPU) happy too.
You are avoiding to tel if it should be 200% better or what to satisfy you? and what those 6 mounts got to do with performance crown of the best performing GPU av the time?

And yet somehow you can declare it a success going off rumors?

25%+ in real world gaming IMO. At nearly 7 months out, we'd be talking about refresh performance, not baseline

[zerazax]

If Nvidia's top card only matches a 5870 then I don't see many price reductions for 5970 coming.

Sadly we won't see any for the 5800s either in that case

[Sam_oslo]

The current 5870 + ~10-20% would cover almost all current games, unless you are using multi-monitor, or some rare demanding games.

Most of us won't need a double-GPU in this round, but yes, in case 5970 would keep it's price until the double-Fermi arrives and beats it. Unless ATi would dare to pushes that over 900MHz then .

[Piotrsama]

The latest PC perspective podcast (~48min mark), Ryan Shrout says "the manufacturability of these chips is pathetic," and the yield is about 10-20% at best.

Somebody should really tell the guys at PcPer to use a video tech that allows to skip to the middle of the video without having to wait for it to buffer those 40+ minutes
Or it's my PC?

[LesGrossman]

You say 20% is disappointing after 6 months, but 20% lead would be the best business scenario for nVidia and make many people (with sans for high-end GPU) happy too.

Going from the rumors, but i'm sorry what lead? There's a thing called 5970. why are you comparing GTX480 with a chip that is almost half the size and have almost half the TDP?
Sucess for nVidia? Right!

Most of us won't need a double-GPU in this round

Off course not! Nvidia can't do it anytime soon why would nvidia fanboys need it

[Macadamia]

Hah, nVidia using Heaven 1.1 (30% more performance in culling, aka Dragon parts of the benchmarks) while ATI uses Heaven 1.0!


It's like Nehalem and Cinebench R10 "R11" dll swap all over again!

[***Deimos***]

As long as AMD uses digital VRM's, what is totaly different game, you just cannot compare it with NVIDIA's VRM.

This only confirms Fermi is really badly made and needs very clear voltage. More capacity is always better, but vendors tend to save costs. They will probably use less caps in order to save few bucks thus I expect (together with rumoured Fermi leakage etc.) very bad OC. At least compared to AMD's clocks targeting 1 GHz.

True that.

AMD HD48xx PCB is so clean and elegant
Likewise HD5870 - although 5970 supposedly have big VRM heat issue

It never ceases to amaze me how far behind nVidia is in terms of PCB design - I find AMD boards simpler and more compact, and with digital VRM, better software control of current/voltage/heat for improved reliability.

At first glance whats the difference how 12V->1.05V. Why not just stepdown buck converter (popular for decades)? this is why
idle/2D__power__cloks
G200's: 60-70W. 300Mhz
4870: 50W. 520Mhz
5870: 20-25W. Clocks drop to 157Mhz! (fast response switch dozens times/sec)
FYI 9600GT is around 30W

Sadly, it is 99.99% likely Fermi will have higher idle power than 5870, and even maybe higher than GT200 (recall the TSMC transistor channel varience power losses)

================================================== ======

about Fermi perfromance - doesn't matter how much faster/slower in 95% of games or anything below 19x12. GF8 is good enough for that.
What matters is very high rez, and very demanding shaders. And with 384 x GDDR5, no bandwidth issue excuse!!

Using Crysis VH 25x16 0AA as "definative" measure for demanding situation:

HD4870 1GB	13
GTX285 1GB	15
HD5850 1GB	18
HD5870 1GB	22
GTX295 1.8GB	24	
HD5970 2GB	28

Also note, HD5870 >> GTX285, but SLI/CF for both about same at 30fps.
30fps on 5870 SLI might be insufficient memory, bandwidth or driver optimization limitation.
Fermi has more SP than GTX295, so it would be big surprise if not faster than HD5870. Hoping for ~30fps.

[thatdude90210]

Somebody should really tell the guys at PcPer to use a video tech that allows to skip to the middle of the video without having to wait for it to buffer those 40+ minutes
Or it's my PC?

Their player is wacky, at least on Firefox for me. Only way to move around is to just download the mp3.

[Chumbucket843]

it will probably have better idle power than gt200. just because peak power increases doesnt mean idle power will. gtx280 @ 230 watts tdp has better idle power than 8800gtx @ 150watts.

you might want to read the article you linked.

We’ve seen Windows Vista-based results that show Nvidia’s GeForce GTX 285 and GTX 260 Core 216 using 20W or so less than ATI’s Radeon HD 4870 1GB. However, in these Windows 7 tests, the Nvidia cards seem to be using as much as 15W more than the older ATI cards.

driver issues. its fixable.

[-Sweeper_]

When tessellation will only have a ON and OFF option and the ON option will enable heavy tessellation that only Nvidia h/w can handle and ATi's 5870 may have a difficult time handling heavy tessellation.

What Nvidia will have is a legit excuse of why the games perform better on Nvidia hardware than ATi one. Just like Nvidia Physx was the keyword for 2009, tessellation may as well be a keyword for Nvidia for 2010 !!

i smell they are already doing that in metro 2033

[Jowy Atreides]

When tessellation will only have a ON and OFF option and the ON option will enable heavy tessellation that only Nvidia h/w can handle and ATi's 5870 may have a difficult time handling heavy tessellation.

What Nvidia will have is a legit excuse of why the games perform better on Nvidia hardware than ATi one. Just like Nvidia Physx was the keyword for 2009, tessellation may as well be a keyword for Nvidia for 2010 !!

If things are as perceived, ATI will have a legit excuse in AA performance.

Fermi shows pretty shocking AA numbers in leaks.

[Macadamia]

i smell they are already doing that in metro 2033

I thought Metro's DX11 was added with Radeons...

Most devs publishing DX11 titles by now have a GTX 400 (repi of DICE was tweeting about it I think) but for Metro I don't think they had time to f*** up the polygonal models that much. Plus, it's probably just using DX11 compute shaders (cough CUDA cough) to speed up post-processing again.

[ElSel10]

I thought Metro's DX11 was added with Radeons...

Most devs publishing DX11 titles by now have a GTX 400 (repi of DICE was tweeting about it I think) but for Metro I don't think they had time to f*** up the polygonal models that much. Plus, it's probably just using DX11 compute shaders (cough CUDA cough) to speed up post-processing again.

Metro 2033 is TWIMTBP

[kgtiger]

I should point out that I really know nothing about this subject.
Just what I have picked up from reading 2/3s of the first fermi thread and all of this one.

While ATI was the first to market with a Dx11 cards by used their old architecture and just adapted Dx11 to it and since fermis new architecture seams to do really well at tessellation.

Will ATI now need to build a new architecture also from the ground up to handle the tessellation more effectively?
Or will ATI be able to hold off building a new architecture from scratch for another generation or two and still get the same tessellation performance that nVidia are getting now.
Or can ATI keep the architecture they have and just keep re-working it?

Another thing I noticed is, if most games are playable at the frame rate that they now have and in the past the best cards have always been judged and rated by how many FPS they could do.
Would it now be fair to say the way cards are now rated might need to change?
If most games don’t need more FPS to improve the smoothness of the game play and tessellation only improves the detail of the actual graphics.
Then maybe we should look at other ways to judge the performance of the new Dx11 cards and take the quality of the graphics into account as well not just the FPS?

With that in mind, from the outside looking in, nVidia does seemed to understand this point some what and as a result have moved there focus to the quality of the graphics more that just the FPS aspect of gaming.
After all is that not what tessellation is all about, improving the graphics quality?
Like the demo with the dragon with tessellation on and the big difference it had with out it, also like the water effect in Just Cause 2.
With my limited experience in gaming to me water has always sucked but that Just Cause 2 demo looked really good.

One other thing, is tessellation the main point of going from Dx10 to Dx11?

Any insight to these question would be much appreciated.