NVIDIA Kepler speculation thread

[DuckThor Evil]

I'm thinking there probably is a fair amount of enthusiasts who will hop from card to card even if it's only marginally faster, for some upgrading in itself is a hobby, buying new things is fun. How many used 7970 do you have to see on sale to make it seem like the market is flooded? Perhaps not that many. The absolutely number can still be quite low.

 

[Albuquerque]

I'm thinking there probably is a fair amount of enthusiasts who will hop from card to card even if it's only marginally faster, for some upgrading in itself is a hobby, buying new things is fun. How many used 7970 do you have to see on sale to make it seem like the market is flooded? Perhaps not that many. The absolutely number can still be quite low.

A casual glance through eBay shows 10 used ones, but I didn't find any on Craigslist. Sooooo, market flooded? Not really, IMO...

 

[nyt]

pricing

FWIW on pixmania.com/fr HD7970 446€ GTX680 499€. Prices going down, good!

 

[DuckThor Evil]

Pixmania has had one model of 7970 for around that price a long time already.

 

[Alexko]

Ooooooh now I like where this could be going, if they were able to implement such a thing.

I find the $ / 99th %tile graph to be useful, but I think you might be able to gain even more insight by one more math step: divide dollars by the 99th %tile avg FPS scale to give you a "Actual Games Perf / Dollar":

Card        Price       FPS         $/Perf (lower = better)
7970        [B]$549[/B]        34          [B]16.15[/B]
680         $499        32          15.59
580         $499        28          17.82
7870        $349        27          12.93
560Ti448    $269        23          11.70

7870 looks quite good here, but so does the 560Ti448 (as expected.)

Fixed.

 

[Albuquerque]

Oh, yeah huh, it's not $599. DERP! Thanks for the correction

 

[Tofu]

If you are running code that is sensitive to DP performance on a Kepler, then you are doing something wrong.

DP is to Kepler as x87 is to Sandy Bridge.

DK

Sorry for OT, but: how would you compute sine, cosine, atan on a Sandy Bridge cpu (and not use x87)?

 

[dkanter]

Sorry for OT, but: how would you compute sine, cosine, atan on a Sandy Bridge cpu (and not use x87)?

Use a math library like MKL.

There are a number of well known techniques for approximating trig functions (e.g. sin x = x for x close to 0).

DK

 

[tunafish]

Sorry for OT, but: how would you compute sine, cosine, atan on a Sandy Bridge cpu (and not use x87)?

With software. The x87 ops are typically actually slower than their SSE software equivalents.

 

[MDolenc]

I know for a fact that the OpenGL driver sees the full text, because that's precisely what happens for iPhone code.

DX has the following call to covert from sorce to byte code:
http://msdn.microsoft.com/en-us/library/windows/desktop/dd607324(v=vs.85).aspx

For OpenGL yes, for DirectX no. DirectX driver doesn't see shader source code. D3DCompile function variants are runtime functions. They have nothing to do with underlying device and are not aware of underlying device. You can call it on a machine without graphics card for all it cares. You tell it shader model and you get byte code. Driver only sees byte code.

 

[Psycho]

DP is to Kepler as x87 is to Sandy Bridge.

What does transcendentals etc have to do with DoublePrecision vs SinglePrecision?

And by kepler I guess you just mean GK104 (I'm sure GK110 will have decent DP performance).

You can call it on a machine without graphics card for all it cares. You tell it shader model and you get byte code. Driver only sees byte code.

...And it's the compiling from byte code to device specific code that takes all the time. And has to be cached in the driver if too slow (AFAIK both amd and nvidia introduced this with BFBC2).

 

[lanek]

I'm actually underwhelmed with both the 7970 and 680. I really would love to see a single GPU pushing out 580SLI or 6970CFX numbers so I could switch to a single card solution without having to give up my current IQ running 3x 19x12 monitors. Based on the benchmarks it doesn't look like 2012 is my year for a single card solution

When i aggree, i think you ask a bit too much there. I use SLI and CFX since the 6600GT ( 6800ultra ) and X1950XTX ), all my configs then have been SLI or CFX. And if there's something i know, if i want a real gain ( performance wise ) over my actual setup, i will forcibly goes for 2 cards then. Ofc there's different factors: When i have switch my 2 good old 5870's for a single 7970 ( before goes ofc for 2 ), performance wise, i was in the equity, but IQ and different factors would allready make my choice goes to the 7970. ( ofc 5870 are olders of 6970-580).

Anyway is someone have seen a review with comparaison in games with FXAA and no FXAA yet or 2.0 vs 3.0 ? Nearly all games tested use FXAA and im a bit curious to see what the difference between FXAA3.0 and 2.0 can be in games.
If games like Skyrim, BF3 etc upgrade to FXAA3.0, this can bring a nice welcomed boost anyway. ( for all brand ofc )

 

[Tofu]

With software. The x87 ops are typically actually slower than their SSE software equivalents.

AFAIK there are no trigonometric functions in MMX/SSE/AVX.

Use a math library like MKL.

There are a number of well known techniques for approximating trig functions (e.g. sin x = x for x close to 0).

DK

MKL would be a good idea - but I need it for a real time system (no external libs) and AFAIK MKL is closed source.

 

[entity279]

He said SSE *software* equivalents

 

[Jawed]

So there's a 16-wide SIMD, a 32-wide SIMD and a 16-wide transcendentals SIMD (and for the sake of completeness a 16-wide load/store unit and the 8-wide DP unit).

Do the 16-wide and 32-wide SIMDs have identical instruction support?

 

[Bob]

So there's a 16-wide SIMD, a 32-wide SIMD and a 16-wide transcendentals SIMD (and for the sake of completeness a 16-wide load/store unit and the 8-wide DP unit).

Do the 16-wide and 32-wide SIMDs have identical instruction support?

They do not (although there is a large overlap). Also, there is no 16-wide SIMD.

 

[fellix]

So there's a 16-wide SIMD, a 32-wide SIMD and a 16-wide transcendentals SIMD (and for the sake of completeness a 16-wide load/store unit and the 8-wide DP unit).

Do the 16-wide and 32-wide SIMDs have identical instruction support?

The only difference, I would think about, is namely the DP support. The vec16 unit could be wired to provide DP math, while the "fatty" vec32 SP unit can be gated off in the mean time. But there's strong insistence of some sources for a dedicated DP, which is not unusual -- remember that GT200 implemented a similar solution, where the DP capability was little more usable than to provide a basic performance for debugging and testing. The same direction, the GK104 is being advertised for.

 

[dnavas]

So there's a 16-wide SIMD, a 32-wide SIMD and a 16-wide transcendentals SIMD (and for the sake of completeness a 16-wide load/store unit and the 8-wide DP unit).

I think it's an 8-wide SFU per scheduler?

 

[trinibwoy]

They do not (although there is a large overlap). Also, there is no 16-wide SIMD.

Amen.

 

[dnavas]

Amen.

And for the slow amongst us, could you spell out what you're talking about?
From what I can tell "A is not B, but they're close. Oh, and there is no B" is a statement that only maps into an escherian slice of reality

-Dave