NVIDIA Kepler speculation thread

[whitetiger]

BSN "confirming"?

I know, but the thing that struck me was the diagram does actually look pretty real - it does look like an nVidia diagram
- a lot of work for someone to put in as a fake....

I wonder how you get 50% more performance with GK110 vs GK104 assuming those specs would even be real, GF114 vs 110 for example with similar difference gives around 30-35% performance difference.

360mm^2 would make it about as big as Tahiti, too

I use some advanced mathematical techniques called arithmetics
- this allows me to calculate that 6 GPCs is a whole 50% more than 4 GPCs
- and in a similar fashion, I calculated than a 384-bit bus is a whole 50% bigger than a 256-bit bus.
- I then calculated that a 512-bit bus is 100% more than a 256-bit bus....
(of course the both the shader clocks & bus clocks will likely end up lower on the bigger part)

Using a further application of this powerful new technique, I found that the GF110 has 33% more SPs than the GF114.
- but the GF114 is clocked much higher, but perhaps is less efficient in it's uses of the SPs due it's 48 SP vs 32 SP architecture

(sorry for the sarcasm, I'm just using the obvious numbers for the obvious reasons)

In the case of the GK104 vs GK110, if the rumors are correct, we're looking at the same 96SP architecture on both... of course that may not be true in reality.... 6 GPCs sounds like an unusual number for nVidia, but we shall see in due course

 

[trinibwoy]

I know, but the thing that struck me was the diagram does actually look pretty real - it does look like an nVidia diagram
- a lot of work for someone to put in as a fake

BSN never claimed that the diagram was from nVidia. It's also very little work to hack up a GF100 diagram and produce that image.

 

[whitetiger]

BSN never claimed that the diagram was from nVidia. It's also very little work to hack up a GF100 diagram and produce that image.

True, but in the end the GK104 will either be 768 SPs with a hot clock, or 1536 SPs and no hot-clock....

Most of the recent rumors point to the latter....

 

[LordEC911]

The GF114 is not just a cut down GF110 so you really can't draw the same conclusions for the GK110 vs GK104.

The same can be said for GK104 vs the big daddy.
So that throws even more variables into the mix.

 

[ninelven]

It's also very little work to hack up a GF100 diagram and produce that image.

Well there is a subtle difference... could still be fake though I suppose...

 

[CarstenS]

BSN never claimed that the diagram was from nVidia. It's also very little work to hack up a GF100 diagram and produce that image.

It's not an Nvidia diagram. Theo never claimed it and even referred to it as their own mockup in his text:
„As you can see in our architectural mockup[…]“

 

[Kaotik]

http://webcache.googleusercontent.c...s/NP700G7C-S01US-features+&cd=1&hl=en&ct=clnk

Samsung accidently confirm GeForce GTX 675M

 

[jaredpace]

http://webcache.googleusercontent.c...s/NP700G7C-S01US-features+&cd=1&hl=en&ct=clnk

Samsung accidently confirm GeForce GTX 675M

Knowing theres a 675m does it mean gk104 is more likely a 680 or a 660?

 

[trinibwoy]

That depends on whether the 675m is based on GK104/GK107/GF114.

From AIDA64:

http://www.aida64.com/downloads/aida64extremebuild1812y4qdz2gtxvzip

GPU information for nVIDIA GeForce GTX 670M (GF114M)
GPU information for nVIDIA GeForce GTX 675M (GF114M)

 

[iMacmatician]

Well if the 600M series Fermis are going up to at least 675M, does that mean mobile Kepler would be 700M series? Also, I'm not sure why NVIDIA would release 40 nm Fermis for the mobile space when 28 nm Keplers are supposedly coming soon (and GK104 doesn't seem far from GF114 in size or power consumption).

That is assuming these are 40 nm. Wikipedia's unsourced 600 and 600M tables lists the GF114M etc. as 28 nm (I don't believe that).

 

[UniversalTruth]

Nvidia’s 28nm mobile lineup leaked

Fermi, not Kepler, as a stopgap

Aug 23, 2011

The short story is this, Nvidia is putting out a bunch of Fermi shrinks on 28nm, and you will likely see the mobile variants first.

 

[Ailuros]

It's February 2012; ask again.

 

[whitetiger]

Given that all the rumors are pointing to a 1536 SP part @950MHz with no hot-clock, giving roughly double the shader performance of the GF114, why would NV only equip it with a 256-bit bus @5GHz - only a 25% improvement on the GF114?


Board price could be an explanation, or pin-limits on the die size
- but AMD managed a 384-bit memory on a similar die size...
- so it seems unnecessarily hampered from a bandwidth point of view

 

[chavvdarrr]

Given that all the rumors are pointing to a 1536 SP part @950MHz with no hot-clock, giving roughly double the shader performance of the GF114, why would NV only equip it with a 256-bit bus @5GHz - only a 25% improvement on the GF114?


Board price could be an explanation, or pin-limits on the die size
- but AMD managed a 384-bit memory on a similar die size...
- so it seems unnecessarily hampered from a bandwidth point of view

Who says die size is same?
Who says bandwidth balance is similar to Tahiti? Any review with detailed review how bandwidth limited is Tahiti in normal situations (eg 1 monitor)?

 

[CarstenS]

Who says bandwidth balance is similar to Tahiti? Any review with detailed review how bandwidth limited is Tahiti in normal situations (eg 1 monitor)?

Here you go buddy:
http://www.computerbase.de/artikel/...deon-hd-7970/20/#abschnitt_384_bit_in_spielen

 

[Ailuros]

Given that all the rumors are pointing to a 1536 SP part @950MHz with no hot-clock, giving roughly double the shader performance of the GF114, why would NV only equip it with a 256-bit bus @5GHz - only a 25% improvement on the GF114?

Probably because it's more important on any architecture how you spend your existing bandwidth than the sterile maximum amount on paper?

Board price could be an explanation, or pin-limits on the die size
- but AMD managed a 384-bit memory on a similar die size...

There still should be a die size difference, whereby Tahiti should still have a significantly higher transistor density per mm2.

- so it seems unnecessarily hampered from a bandwidth point of view

That's something we still have to find out. Initial reasoning is obviously in that direction. But in order to get there we'd need deeper architectural details instead of just a couple of sterile unit amounts.

 

[CarstenS]

Significant higher transistor density while at the same having a larger IO-part (display controller, memory controller) sounds like AMD must have way more highly squeezable SRAM on die - which is quite counter-intuitive to GK104s purported insenstivie to 50% lesser bandwidth. I'm quite puzzled.

 

[Ailuros]

You're not alone. It's the typical point where you'd think you're close enough to real specifications and still don't know much more than before.

 

[whitetiger]

Who says die size is same?
Who says bandwidth balance is similar to Tahiti? Any review with detailed review how bandwidth limited is Tahiti in normal situations (eg 1 monitor)?

I was primarily comparing it to the GF114
- if I assume your spec that you gave a few pages back are close enough (and all rumors are saying the pretty much the same thing, so let's just go with that, for the sake of arguement)
- then the GK104 has 2x the Shader performance of the GF114, but only 25% more bandwidth.
- so unless the GF114 had an excess of bandwidth in common scenarios, then this would suggest that the GK104 is going to tend to be somewhat starved of bandwidth...

Perhaps one could argue that the GK104 is actually very close to Cayman in terms of SP performance - same number shaders, and similar clocks
- but I think the NV scalar architecture is generally known to be 'more efficient' than the VLIW4/5 architectures - ok, not so much 'more efficient' as more performance per apparent MFLOP.
- so given that, the GK104 would still appear to be under-endowed on the bandwidth front
- I'm not the first one to say this, and not the last probably either.

I think you missed my point about the die sizes
- my point was this - the Tahiti and GK104 die sizes are known/believed to be similar (the Tahiti is/thought to be a little bigger)
- one reason for not including a 384-bit bus on the GK104 die could be pin-out limitations
- ASICs are often pin-limited, meaning they can't fit more than a certain number of pins on a certain die size....
- if the GK104 is similar in size to the Tahiti, then it probably would have been feasible to have a 384-bit bus on the die, if NV felt it was necessary (or 320bit like the GTX570)

- so, purely speculating here, if the GK104 is bandwidth limited, then NV could have conceivably put a 384-bit (or 320-bit) bus on the chip, but are keeping that back for now, in order to hit a better price point with the first batch of board.

Well, I know, that is pure speculation (which is ok in a Kepler speculation thread)

Otherwise, it does seem odd to produce a chip that it so biased towards shader performance (assume that the GF114 was 'balanced')

 

[Ailuros]

In what way was GF114 exactly "balanced" compared to GF110 when it comes to aspects like raw GFLOPs on paper and texel fillrate compared to each chips bandwidth?