NVIDIA GF100 & Friends speculation

[Lonbjerg]

Wall street suits won't be too happy if you win benches by 10% but cost, say 50% more to produce.

And yes, they do matter, no matter how you feel about them.

That is the same BS as the GT200 vs RV770...and NVIDIA made a healthy profit in that timeperiod.
Broken record is broken record.

 

[hatter]

The closer it gets, the warmer the air.

It gets there fast, for what it's worth.

Hmmm... is an overclocked GTX480 matching HD 5970? Difficult to guess from leaks, hints so far but it's silly season after all

 

[rpg.314]

That is the same BS as the GT200 vs RV770...and NVIDIA made a healthy profit in that timeperiod.

May be you should look at nv's margins before and after rv770 for a better clue. AMD's margins in 2Q10 will be even more illustrative.

 

[Picao84]

What about another type of efficiency?

We are always talking about perf/watt, perf/area or perf/transistor count, and truthfully, i agree.

But something is striking me as odd, since the first rumours about GF100. It seems that GF100 is not much powerfull in terms of raw, FLOPS power, not passing the 1,5 TFLOPS barrier. However, it still manages to be at least equal to HD5870 in performance, a part which has 2,7 TFLOPS of power. So, in this sence it would win handily in somekind of perf/flops measures.

Could we extrapolate from here, that ATI, without changing their architecture, would be in the future on a less brightly spot? I mean, Cypress seems to have need for much much higher flop power than GF100. I wonder if a GF100 derivate at 22nm tech would really be a beast.

 

[fbomber]

The thing is, at the time, I don't think it was actually true, not like it is today. Compare, for instance, the GeForce 4 vs. Radeon 8500. The GeForce 4 was a refresh of the GeForce 3 architecture, while the Radeon 8500 was a brand-new architecture that came out within a few months of the GeForce 4 parts. Yes, the 8500 was more feature rich, but this is more because of the fact that it was released over a year later than due to any sort of difference in design decisions (although there were differences in design decisions, of course).

Then again if you look at the Radeon 9700 and its derivatives vs. the GeForce FX and derivatives, well, the FX just didn't perform, so it definitely didn't fit this.

And when you look at the GeForce 6x00, when that first came out it was still competing with ATI's own 9700 derivatives, and so it was quite a lot more feature-rich (and also higher-performing). Later, when ATI released their own SM3 parts, obviously nVidia lost this advantage.

So basically, nVidia and ATI have been enough out of sync that we never really saw much of any "raw power vs. features" competition directly. Instead they leapfrogged one another in both features and performance.

This time around, by contrast, the GF100 should have been released at almost the same time as ATI's new architecture. It ended up a bit later, of course, but we can be pretty darned sure that no new features have been added to the GF100 in the mean time. So this is a true competition in design strategies, and those design strategies are quite transparent: ATI has gone for a much larger number of less flexible units than has nVidia. They have, in effect, gone for much higher raw performance (even if said performance isn't realized in real-world situations), while nVidia has gone for lower raw performance but instead sought to get better performance in the end through better use of the available compute power.

Please bear in mind that I never meant that nVidia has actually shot for a lower performance target. That wasn't my intention at all: I'm sure nVidia is every bit as interested in attaining the performance crown as ATI. I'm just saying that their way of going about it has been rather different.

No. 8500 was released much earlier than geforce 4. In fact, Nvidia released the geforce 3 Ti500 to compete with the 8500.

 

[Mintmaster]

I don't think luck had anything to do with any of this. It's about making a series of design decisions. Obviously ATI went for raw power, while nVidia went for a more flexible, efficient design capable of more different things.

How is nVidia's design more efficient when it needs more silicon than ATI to achieve the same performance? ATI was more flexible when it came to AA control in the pixel shader, hence DX10.1 support. The only area they lagged in was local store and GPGPU software.

NVidia's most efficient design by far since DX10 came around was G92/G92b. It took a near-miraculous engineering effort in RV770 to leapfrog it. Since then, NVidia has been regressing on that front.

So basically, nVidia and ATI have been enough out of sync that we never really saw much of any "raw power vs. features" competition directly. Instead they leapfrogged one another in both features and performance.

I still see a shift in NVidia's philosophy from G80 onwards. Before that, NVidia would go for perf/mm2 above all else, adding just enough silicon for features to get the checkbox. ATI would do it right when they added features, often to their own demise because it wasn't used enough.

Now they're both adding well thought-out features, except NVidia is trying to show that it's better at GPGPU. Despite the radical architectural changes, though, it doesn't look like a cut down, 330mm2 Fermi would be generally faster than RV870 even at that.

Let me just add in a small comment that if nVidia had no problem surviving the GeForce FX era business-wise, they'll have no problem now, especially as the GF100 looks like it will, at the very least, be a vastly better part compared to the competition than the FX was.

You have to realize that the reason NVidia survived that era is that it had a DX9 budget part in the FX5200, and the OEMs gobbled it up. It didn't matter that the Radeon 9000 smacked it silly, because it didn't have the DX9 checkbox.

This time, NVidia is going to get crushed in the mainstream and budget markets. NVidia has even said that it plans to use its DX10.1 parts in this market for a while, and the GF1xx parts are going to offer even worse value to get that DX11 checkmark.

 

[Rangers]

We are always talking about perf/watt, perf/area or perf/transistor count, and truthfully, i agree.

But something is striking me as odd, since the first rumours about GF100. It seems that GF100 is not much powerfull in terms of raw, FLOPS power, not passing the 1,5 TFLOPS barrier. However, it still manages to be at least equal to HD5870 in performance, a part which has 2,7 TFLOPS of power. So, in this sence it would win handily in somekind of perf/flops measures.

Could we extrapolate from here, that ATI, without changing their architecture, would be in the future on a less brightly spot? I mean, Cypress seems to have need for much much higher flop power than GF100. I wonder if a GF100 derivate at 22nm tech would really be a beast.

Flop power either way isn't really relevant though, it just isn't. It boils down to perf/mm.

 

[neliz]

Cypress seems to have need for much much higher flop power than GF100.

"a need for higher flop power"? What is that?

Do you mean that Evergreen's architecture could allow for a higher computational throughput if specifically coded for it? But in essence still provides the same performance as GF100 in an unoptimized situation?

I'm not sure how you're trying to spin Evergreen's higher TFLOP numbers as something negative. Trying to divide a game framerate by a theoretical number is borderline insane, that's not even something nVidia would use in its presentations and guides

OpenCL applications have already proven to get pretty good utilization out of Evergreen, even at 80% efficiency, a GF100 at 95% efficiency would still lose to Cypress.

 

[Razor1]

We are always talking about perf/watt, perf/area or perf/transistor count, and truthfully, i agree.

But something is striking me as odd, since the first rumours about GF100. It seems that GF100 is not much powerfull in terms of raw, FLOPS power, not passing the 1,5 TFLOPS barrier. However, it still manages to be at least equal to HD5870 in performance, a part which has 2,7 TFLOPS of power. So, in this sence it would win handily in somekind of perf/flops measures.

Could we extrapolate from here, that ATI, without changing their architecture, would be in the future on a less brightly spot? I mean, Cypress seems to have need for much much higher flop power than GF100. I wonder if a GF100 derivate at 22nm tech would really be a beast.

This is a design choice, even though nV's chips have half the theoretical flops it performs competitively and surpasses ATi's counterparts, but die size is the cost of that efficiency. So at the end of it all, we have to see what that efficiency does for GPGPU applications (as for games we see that) vs. ATi's cards when there are comparable real world applications, not theoretical. If the trade off doesn't pay off (die size and efficiency vs. raw flops) well then its a wash, if it does then its got alot of merit.

 

[Picao84]

NVidia has even said that it plans to use its DX10.1 parts in this market for a while, and the GF1xx parts are going to offer even worse value to get that DX11 checkmark.

Hmm.. They did tape out GF108 some days ago, it seems. nVIDIA code numbers ending with "8", are usually low end parts.

So, i dont think that slip up was really meaning it. It would be bad move if he would say: "our current parts on the market are lame ducks. We will have something shortly. Stay tuned". Then noone would buy them. Its just PR. Its not to be taken seriously

 

[Mintmaster]

So, in this sence it would win handily in somekind of perf/flops measures.

It's a useless metric.

It's like a fat guy saying he's a pretty good weightlifter for someone with 30% bodyfat. Unfortunately, he gets killed by people in his weight class. Maybe he can lift as much as someone with more muscle, but that guy is in a lower weight class.

 

[Groo The Wanderer]

I doubt 480sp vs 512sp would make a lot difference, but the clocks (600 core vs 700, and 700 mem vs 900) sure would make a big difference.

BTW whoever that guy is I'm 100% certain that he was the source for Charlie's "5% perf with 600mhz" remarks.

Nope, not even close.If you read my article, you would understand why.

-Charlie

 

[Picao84]

I'm not sure how you're trying to spin Evergreen's higher TFLOP numbers as something negative.

Dont get me wrong, im not trying to spin anything. If anything, ATI does it by promoting FLOPS on their gaming PR, knowing it has no meaning to graphics

But, i ask from you: Is it conceivable to imagine an HD5870 with less FLOPS and same performance? If answer is yes, ok. If answer is no, then my first sentence stays: Cypress needs more FLOPS power to perform.

And then, here we are saying GF100 is mainly directed at GPGPU, when with less FLOPS it does more graphically

 

[seahawk]

Dont get me wrong, im not trying to spin anything. If anything, ATI does it by promoting FLOPS on their gaming PR, knowing it has no meaning to graphics

But, i ask from you: Is it conceivable to imagine an HD5870 with less FLOPS and same performance? If answer is yes, ok. If answer is no, then my first sentence stays: Cypress needs more FLOPS power to perform.

And then, here we are saying GF100 is mainly directed at GPGPU, when with less FLOPS it does more graphically

And? If you look at the results ATI builts a smaller GPU with more Flops and real hardware tesselation that perfroams equally to the huge monster NV made, and which they could not produce for 6 months after 580X0 came out.

Cypress does not need more Flops, it simply has more Flops.

 

[Razor1]

you can drop the real hardware tessellation stuff

 

[Xentropy]

Poster here claims that Nvidia bought a license for the Heaven benchmark and produced version 1.1 themselves:

http://www.overclock.net/8840643-post67.html

We never got to get out hands on 1.1, and Unigine would not send it to me, stating that they did not produce it. I later found out that nVidia bought a license and produced 1.1. As far as I can see, this is not a real world scenario.

 

[KimB]

In HPC, the cpu *has* to run x86. This discussion isn't about HTPC, but HPC.

Sorry, meant HPTC (high-performance technical computing), which is more or less another word for HPC. And no, I strongly disagree with this claim. It might be valid in areas where Windows is popular, but most other operating systems have no trouble whatsoever running on other architectures. I'm pretty sure that Linux is the OS of choice in most HPC environments.

 

[Picao84]

And? If you look at the results ATI builts a smaller GPU with more Flops and real hardware tesselation that perfroams equally to the huge monster NV made, and which they could not produce for 6 months after 580X0 came out.

Cypress does not need more Flops, it simply has more Flops.

If you read my original post, im not concerned with Cypress. Cypress is a great GPU and ATI deserves kudos for it. I was talking the future of that architecture. If there would not be diminishing returns for a given architecture, noone would ever change them, with the constantly improving miniaturization (probably wrongly spelled) technologies (90nm -> 80 nm -> 65 nm....). Anyway, sorry for provoking an off-topic trainwreck.. This is my last answer to this...

 

[rpg.314]

But, i ask from you: Is it conceivable to imagine an HD5870 with less FLOPS and same performance?

Certainly, if you re-balance the chip by improving other parts. Witness 480, ~half the flops and better performance.

 

[Mintmaster]

Not really. GF100 vs Cypress is a better situation for NV then GT200 vs RV770 was. They now have a lead in features, they have a clearly more future-proof products

This is completely untrue. GT200 vs RV770 was a much better situation for NVidia. GT200 had an advantage in idle power consumption, much better local memory usability, no competition with CUDA, and could use its efficient previous gen parts (G92b, G94b, etc) in the low end without any consequence, and was available shortly before RV770.

Now, Evergreen is blazing fast at atomics, has good local memory support, has DirectCompute and OpenCL to use in GPGPU, has a huge feature advantage in the low end (DX11 vs DX10.1), lower power consumption, has been available for a long time at all price points, and the only real disadvantage it has is speed in certain games at the $500 price point.