NVIDIA Kepler speculation thread

[AnarchX]

GF114 680/670M
GK107 650/640M
GF118/119 630M etc.. later when GK106 is ready nV may launch it as 685M and 675M

GF117, which could replace GeForce 610M, GT 620M and GT 630M.
Latest GeForce 295.55 also lists GF117-SKUs.

It seems a kind of funny that GK107 is crippled by DDR3, while GF117 has to use GDDR5 to reach adequate BW on its 64-Bit IMC.

GTX 660M @ Cebit:

http://www.4gamer.net/games/120/G012093/20120306077/

And some kind of interview with Nick Stan from Nvidia:

"Users of our products from traditional" performance only "about 1.5 times more information. Kepler would not be satisfied, but not obvious, and performance per unit of power consumption is up significantly, and conventional products of the same price range should also significantly improve performance compared to the absolute. "

... and more about Hyper-Threading CUDA cores?

 

[Picao84]

... and more about Hyper-Threading CUDA cores?

Holy shit! I did speculate about Hyper-Threading CUDA cores today on ABT's forum :O

I also have my doubts.. 1536 cores @1.4Ghz? 1536 cores with hotclock in such a small chip? Or maybe NVIDIA invented something like Hyperthreading, with 768 cores behaving almost like 1536, and thus the system actually reports the latter? Just speculating here, because I could see 768 core with hotclock on such a small chip, but having a hard time coping with 1536

http://alienbabeltech.com/abt/viewtopic.php?p=60140#p60140

 

[AnarchX]

GPUs are multi-threaded since years.

Maybe here is some native Japanese speakers, who can translate it...

 

[Picao84]

GPUs are multi-threaded since years.

Maybe here is some native Japanese speakers, who can translate it...

Yes, but do they work the same as Intel Hyperthreading? Honest question, I have no idea. It doesnt make any sense at all to talk about a Hyperthreaded CUDA core?
There was also that guy on SemiAccurate forum saying GK104 has only 512 cores but which somehows behaves like 1536... Maybe he wasnt far from the truth after all?

 

[Ailuros]

Yes, but do they work the same as Intel Hyperthreading? Honest question, I have no idea. It doesnt make any sense at all to talk about a Hyperthreaded CUDA core?
There was also that guy on SemiAccurate forum saying GK104 has only 512 cores but which somehows behaves like 1536... Maybe he wasnt far from the truth after all?

How about a much more reasonable scenario? Have the GPU run at default for N Watts TDP and with a sw switch for example allow it to clock higher for a N+ Watts TDP?

 

[AnarchX]

But 1,4GHz for ALU+TMU+ROP, should be hardly reached with air cooling.

 

[Picao84]

How about a much more reasonable scenario? Have the GPU run at default for N Watts TDP and with a sw switch for example allow it to clock higher for a N+ Watts TDP?

How would that translate into Kepler's supposed energy efficiency?

 

[chavvdarrr]

How about a much more reasonable scenario? Have the GPU run at default for N Watts TDP and with a sw switch for example allow it to clock higher for a N+ Watts TDP?

cheater

 

[Gipsel]

How would that translate into Kepler's supposed energy efficiency?

It's quite efficient at stock values?

 

[Ailuros]

How would that translate into Kepler's supposed energy efficiency?

A real continious hotclock as up to now would be IMHO a horrible idea with a very tight transistor density. What I suggested wasn't a hotclock or anything close to twice the core frequency. Just allow the GPU to clock slightly higher at the cost of higher power consumption.

If you start out at default with a low enough power consumption, it doesn't necessarily break any perf/W ratio.

 

[A1xLLcqAgt0qc2RyMz0y]

http://www.fudzilla.com/home/item/26179-kepler-256-bit-faster-in-dx11-games-than-7970

Now it's only faster in DX11. Next thing you know it'll only be faster on weekends.

Previously it was only known to be faster than the 7950 and sometimes faster than the 7970.

Now it is known to be faster than the 7970 in all DX11 games.

This is an improvement from what was known before.

 

[Picao84]

A real continious hotclock as up to now would be IMHO a horrible idea with a very tight transistor density. What I suggested wasn't a hotclock or anything close to twice the core frequency. Just allow the GPU to clock slightly higher at the cost of higher power consumption.

If you start out at default with a low enough power consumption, it doesn't necessarily break any perf/W ratio.

Oh ok, so you stay by the stance that there is really no hotclock. I was talking in the context of there still existing an hotclock, like the latest rumours suggest.

 

[Ailuros]

Oh ok, so you stay by the stance that there is really no hotclock. I was talking in the context of there still existing an hotclock, like the latest rumours suggest.

Entertain me (and yes it's an honest question): where does that rumor come from exactly? Site? Author?

Previously it was only known to be faster than the 7950 and sometimes faster than the 7970.

Now it is known to be faster than the 7970 in all DX11 games.

This is an improvement from what was known before.

Sarcasm? Cause it's really the same thing just worded differently.

 

[Picao84]

Entertain me (and yes it's an honest question): where does that rumor come from exactly? Site? Author?



Sarcasm? Cause it's really the same thing just worded differently.

VR-Zone.
Not that Im defending it. I was just playing by it, trying to imagine how it could be true

 

[Ailuros]

VR-Zone.
Not that Im defending it. I was just playing by it, trying to imagine how it could be true

ok.....

 

[Sxotty]

Entertain me (and yes it's an honest question): where does that rumor come from exactly? Site? Author?



Sarcasm? Cause it's really the same thing just worded differently.

Those are not the same at all.

"Faster than X, sometimes faster than Y"
Could mean that it was only faster in one specific side case.
"Faster than X and faster than Y in directX11"
Means it is faster in almost all new games. Those are very different things. I actually find it unlikely to be true given then even when one was clearly superior there was usually a weird game out there that showed the opposite.

 

[Ailuros]

Those are not the same at all.

"Faster than X, sometimes faster than Y"
Could mean that it was only faster in one specific side case.
"Faster than X and faster than Y in directX11"
Means it is faster in almost all new games. Those are very different things. I actually find it unlikely to be true given then even when one was clearly superior there was usually a weird game out there that showed the opposite.

Since the games/applications one could label as "DX11" are relatively few, that previous "sometimes faster than 7970" doesn't seem to make any significant difference.

Here's the exact wording at fudzilla, which is not necessarily accurate:

The GK104 is much smaller than Tahiti, Radeon HD 7970 family and according to our intel it should end up faster in some cases. In all DirectX 11 games, Nvidia’s GK104 based Kepler card should end up faster, at least this is what a few sources are confirming.

http://www.fudzilla.com/home/item/26179-kepler-256-bit-faster-in-dx11-games-than-7970

 

[jimbo75]

We could try a small leap of faith and equate "all dx11 games" Fudo style to actually mean "all games with heavy tessellation". Suddenly it all seems reasonable.

 

[Ailuros]

We could try a small leap of faith and equate "all dx11 games" Fudo style to actually mean "all games with heavy tessellation". Suddenly it all seems reasonable.

Overanalyzing tidbits like that is a waste of bytes anyway. But since fudzilla was also quick to "adopt" VR-zone's hotclock claim: 1536SPs * 2 FLOPs * 1,411GHz = 4,335 TFLOPs

 

[dnavas]

Overanalyzing tidbits like that is a waste of bytes anyway. But since fudzilla was also quick to "adopt" VR-zone's hotclock claim: 1536SPs * 2 FLOPs * 1,411GHz = 4,335 TFLOPs

That does seem silly. Is MAD no longer single-clocked, but any pair of ops can use result of first as input of second? One op-pair per non-hot-clock? Would reduce expense of transporting operands around and might result in efficiency gains (suppose it depends on how many pairs could be usefully constructed). Was there a VLIW rumor at some point?