NVIDIA Kepler speculation thread

[UniversalTruth]

If GPU A has twice the perf/watt than B, than, at any given targeted power consumption (it may be a little simplistic), A is going to be twice as fast as B. In this case, 100W.

Very doubtful to me, I mean a given architecture can scale differently, so at half the TDP, this same chip can provide more or less performance.
Perf/watt can have different values for one and the same chip at different ends of the consumption spectrum.

 

[KimB]

Very doubtful to me, I mean a given architecture can scale differently, so at half the TDP, this same chip can provide more or less performance.
Perf/watt can have different values for one and the same chip at different ends of the consumption spectrum.

While true, the discrepancy is rarely this drastic. What usually happens in practice is that one architecture tends to have a lead in performance per watt pretty much across the board, but the size of that lead tends to vary depending upon the specific product.

 

[mczak]

While true, the discrepancy is rarely this drastic. What usually happens in practice is that one architecture tends to have a lead in performance per watt pretty much across the board, but the size of that lead tends to vary depending upon the specific product.

I think you've got the right idea here, but I'm not sure this still holds true with dynamic clocked products - for example if you have an app which causes your units to be underutilized this means it can turbo up, and while that means performance goes up a bit unfortunately power efficiency will go down (well, unless you could clock it up without raising voltage).

 

[AnarchX]

Acer G3620 with GTX 660 1.5GiB available on 25th June: http://www.elkjop.no/product/data/stasjonar-pc/ASG3620002/acer-stasjonar-pc-aspire-predator-g3620

 

[lanek]

1.5gb? so i can assume a 192bits memory bus ( i dont see them going for 384bits )

 

[LordEC911]

1.5gb? so i can assume a 192bits memory bus ( i dont see them going for 384bits )

A cutdown GK104?

 

[AnarchX]

Presumable there could be a difference between OEM and retail 660 (see 560: OEM=GF110, retail=GF114).

 

[A1xLLcqAgt0qc2RyMz0y]

A cutdown GK104?

No the GK106.

http://www.techpowerup.com/163889/NVIDIA-GK106-GPU-Detailed.html

Its specifications are listed below.

• 28 nm, around 210 mm² die-area, Kepler architecture
• Two Graphics Processing Clusters (GPCs), four Streaming Multiprocessors (SMXs)
• 768 CUDA cores
• 64 TMUs, 24 ROPs
• 192-bit wide GDDR5 memory interface, holding 1.5 GB or 2 GB memory
• PCI-Express 3.0 bus interface
• Around 130W TDP

 

[Gipsel]

No the GK106.

http://www.techpowerup.com/163889/NVIDIA-GK106-GPU-Detailed.html

Its specifications are listed below.

• 28 nm, around 210 mm² die-area, Kepler architecture
• Two Graphics Processing Clusters (GPCs), four Streaming Multiprocessors (SMXs)
• 768 CUDA cores
• 64 TMUs, 24 ROPs
• 192-bit wide GDDR5 memory interface, holding 1.5 GB or 2 GB memory
• PCI-Express 3.0 bus interface
• Around 130W TDP

So virtually the same die size as a Pitcairn (iirc 212mm²). Appears a bit large compared to GK104 (71% of GK104's die size) considering it has only half the front end/shaders/TMUs and three quarters of the ROPs and memory interface. I guess it will have a hard time competing with Pitcairn on performance.

 

[Homeles]

No the GK106.

http://www.techpowerup.com/163889/NVIDIA-GK106-GPU-Detailed.html

Its specifications are listed below.

• 28 nm, around 210 mm² die-area, Kepler architecture
• Two Graphics Processing Clusters (GPCs), four Streaming Multiprocessors (SMXs)
• 768 CUDA cores
• 64 TMUs, 24 ROPs
• 192-bit wide GDDR5 memory interface, holding 1.5 GB or 2 GB memory
• PCI-Express 3.0 bus interface
• Around 130W TDP

It's also possible that it's a further cut-down GK104. There's currently no evidence supporting either scenario though.

 

[trinibwoy]

Even with all these devastating yield and availability issues the 680 outsold the 7970 by almost 3:1 (+0.25% vs +0.09%) last month according to the latest steam survey. Must be magic fairies.

 

[Homeles]

Even with all these devastating yield and availability issues the 680 outsold the 7970 by almost 3:1 (+0.25% vs +0.09%) last month according to the latest steam survey. Must be magic fairies.

While Steam surveys are hardly a reliable indicator for anything other than trends over a long period of time, it is pretty clear that GTX680 is simply selling well.

 

[trinibwoy]

While Steam surveys are hardly a reliable indicator for anything other than trends over a long period of time, it is pretty clear that GTX680 is simply selling well.

Show me a more reliable indicator and I'll gladly use it If a survey with millions of samples is unreliable then I fear for any other statistic out there.

 

[DSC]

http://www.geforce.com/whats-new/ar...-next-gen-gtx-680-powered-real-time-graphics/

Interview with Tim Sweeney about UE4.

 

[Ernestds]

Show me a more reliable indicator and I'll gladly use it If a survey with millions of samples is unreliable then I fear for any other statistic out there.

There is none. But ffs stop using Steam Survey. It is a mess of numbers that would never make sense in real world (with real consumers).

And no, I am not saying anyone is not outselling anyone.

 

[silent_guy]

Are there reasons to not trust Steam for comparative uses for products that are sold at the same time and that target the same segment? Not speaking about absolute numbers in general, because participation rate in the survey may be very different between segments, but within the constraints set above, is it likely that, say, an AMD user will be less inclined to submit results than an Nvidia user?

 

[Otto Dafe]

Are there reasons to not trust Steam for comparative uses for products that are sold at the same and that target the same segment? Not speaking about absolute numbers in general, because participation rate in the survey may be very different between segments, but within the constraints set above, is it likely that, say, an AMD user will be less inclined to submit results than an Nvidia user?

Does either company give away Steam keys / Steamworks games with their boards? That could skew things rather drastically.

 

[Ryan Smith]

Does either company give away Steam keys / Steamworks games with their boards? That could skew things rather drastically.

AMD does. Their various game deals are almost always through Steam. DiRT: Showdown for example is being given away with most new Radeons right now.

NVIDIA on the other hand doesn't usually give games away. IIRC a lot of their partners were giving away BF3 last year, but I can't recall the last time anyone was giving away Steam games.

Edit: Whoops, forgot about Batman: Arkham City last year. NVIDIA was giving that away and it was being distributed over Steam. However that deal ended well before the 600 series launched.

 

[Alexko]

Does either company give away Steam keys / Steamworks games with their boards? That could skew things rather drastically.

AMD gives away Steam keys for DIRT, maybe another game or two; perhaps NVIDIA does the same thing on occasion, I don't know.

All in all I doubt it has a very significant effect on the results, which are definitely skewed towards the high-end, but I don't think there's much brand bias.

So unless some solid, concrete information surfaces, I think we can put the terrible yields rumors to rest.

 

[trinibwoy]

But ffs stop using Steam Survey. It is a mess of numbers that would never make sense in real world (with real consumers).

What could possibly make you think that the picture Steam paints is very different from reality? Steam is as real world and real consumers as you get when it comes to consumers of discrete graphics cards - i.e. people who purchase and play 3D PC games.

There can be up to 5 million people logged into Steam at any given instant. That's not representative of the real world? Then what is?

The impact of Steam vouchers is probably negligible. Steam is so massive and popular now that anyone who gets a voucher for a game is probably already on Steam anyway. All the people poo-pooing Steam's numbers need to stop blowing hot air and come up with actual reasons for why it's not representative.

is it likely that, say, an AMD user will be less inclined to submit results than an Nvidia user?

Nope. Same goes for a 7970 owner vs a 680 owner etc. There's no selection bias in play here and the sample size is massive. Pretty damn good proxy if I've ever seen one.