Nvidia GT300 core: Speculation

[dnavas]

Modern workloads will be more and more about being fast at less regular computations

Hmm, are you arguing for moving some of the "SF" instructions into the SPs? I would think log/rcp would be really useful there (hence my previous link). SF could be relegated to sin/cos approximations, or those blue dots might be something else entirely.

and data structures

...caches?

Just trying to keep up.

 

[nAo]

No, I am not arguing for that.

 

[w0mbat]

Performance figures provided along with the specifications talk about 70% of a 5870CF setup. So 20% over a 5870.
Hence the notion of a X2, it is required for performance leadership.

That would be a hughe, negative, surprise. I see GF100 ahead or minimum on par with HD5870 X2.

 

[ShaidarHaran]

That would be a hughe, negative, surprise. I see GF100 ahead or minimum on par with HD5870 X2.

Seems unlikely to me. Such a chip would need to be well over twice as fast as GT200. On paper, sure, but in real world apps, I don't think so.

 

[w0mbat]

I think 2.5*GT200 was the main goal.

 

[neliz]

That would be a hughe, negative, surprise. I see GF100 ahead or minimum on par with HD5870 X2.

KonKorT has his doubts too
http://www.hardware-infos.com/news.php?news=3222

There are other reasonable doubt that the Nvidia chip purely "perform AMD's dual solution is able to keep up. Sicher scheint jedoch, dass der Abstand kleiner ausfallen wird als noch zwischen HD 4870 X2 und GTX 285 und dass der GT300 im direkten Vergleich deutlich weniger Energie benötigen wird. Sure seems that the gap is smaller than it was canceled between HD 4870 X2 and GTX 285 and that the GT300 will require significantly less energy in a direct comparison.

 

[Razor1]

well that would be pretty impressive if they do that.

 

[flynn]

that is not what i meant

i believe we will see much more improvement in parallel processing; Nvidia's direction away from pure gaming

How big is the market for that? I know about DirectCompute but I'm wondering how much is Microsoft is willing to invest on that to have software vendors adopt it. I'd rather they had adopted OpenCL instead of pushing their proprietary and incompatible version, though.

 

[ShaidarHaran]

I think 2.5*GT200 was the main goal.

I've heard that number as well, but honestly no one outside the design team really knows if that was the internal goal. Also, 2.5x specs doesn't lead to 2.5x performance gain, as RV770->RV870 has shown us.

 

[Razor1]

not the same thing as doubling units the gf100 is different.

Is it me or is anyone else see that the forums are going very slow when posting?

 

[Arty]

Well anything short of CypressX2 would be disappointing, according to the same folks who are 'dissappointed' by Cypress.

Personally anything over 40% (average) faster than Cypress makes me feel funny inside.

 

[MfA]

How big is the market for that?

Also, what does it do for me ... I don't need office to run any faster and I don't do video encoding ... so unless it can make porn look a whole lot better the only thing it can do for me is run games.

 

[jaredpace]

Anyone have a guess at the GF100 die area? If transistor density increases by 1.75x (as for ati going 55nm -> 40nm), then nv's 3.0bln 40nm tran = 575mm^2 and 3.2bln = 615mm^2.

 

[dnavas]

No, I am not arguing for that.

So by "less regular computations" you were referring not to, specifically, instruction set (which your reference to MADD seemed to indicate), but rather to the diversity of problems that need to be adequately tackled (MADD being pretty specifically targeted, and hence a flop rating that relied on it being increasingly unuseful)?

I am curious where you think the instruction architecture is going, though. Dedicated simple ALUs with VLIW front-ends, or fewer, more complicated ALUs? Vector/SIMD vs. MIMD? Fixed/managed on-chip memory, or relatively flexible, coherent caches?

 

[MfA]

They are probably both already VLIW+SIMD at this point (well NVIDIA is more LIW+SIMD but same difference). Whatever else happens VLIW is there to stay for a while yet IMO.

 

[Ailuros]

I've heard that number as well, but honestly no one outside the design team really knows if that was the internal goal. Also, 2.5x specs doesn't lead to 2.5x performance gain, as RV770->RV870 has shown us.

That's absolutely true; however the deeper the architectural changes the more the chances for achieving higher efficiency.

 

[DegustatoR]

that is not what i meant

i believe we will see much more improvement in parallel processing; Nvidia's direction away from pure gaming

I understand. But it looks to me like that "pure gaming" you're speaking about is moving towards parallel processing fast.

That would be a hughe, negative, surprise. I see GF100 ahead or minimum on par with HD5870 X2.

Why would a chip with +50% complexity be "minimum on par" with dual HD5870?..
If anything it should be approximately +50% to Cypress and Hemlock will probably be around +70% to Cypress. Of course it can end up being +20% as GT200 compared to RV770 did or +70% which will put it against Hemlock. But I don't see any reason to expect Hemlock performance level as a minimum.

 

[seahawk]

-20% on average in the typical benchmarks to the competitions best X2 card is a clear victory as it would dominate the X2 in the worst case scenarios. Getting a draw to the X2 would be domination.

 

[KonKort]

KonKorT has his doubts too
http://www.hardware-infos.com/news.php?news=3222

Yes, I have got my doubts, but as I say: The performance difference between Geforce 380 and Radeon HD 5870 X2 will be much smaller than Geforce GTX 285 and Radeon HD 4870 X2 and I will not exclude that Geforce 380 is even faster.
Let's look to the worst case: HD 5870 X2 is 20% faster than Geforce 380. Then you must ask you for which price, because Geforce 380 is a Single-GPU (no Multi-core profiles, no micro stuttering etc.), who will not consume more energy than GTX 280. If I look to HD 5870 X2, I hope it will consume under 275 watts.

 

[Kaotik]

Yes, I have got my doubts, but as I say: The performance difference between Geforce 380 and Radeon HD 5870 X2 will be much smaller than Geforce GTX 285 and Radeon HD 4870 X2 and I will not exclude that Geforce 380 is even faster.
Let's look to the worst case: HD 5870 X2 is 20% faster than Geforce 380. Then you must ask you for which price, because Geforce 380 is a Single-GPU (no Multi-core profiles, no micro stuttering etc.), who will not consume more energy than GTX 280. If I look to HD 5870 X2, I hope it will consume under 275 watts.

Worst case? For all we know, it could be another NV30-case, the worst case is much, much worse than that