NVIDIA Kepler speculation thread

What I don't get is that with the death of the budget market due to integrated graphics approaching their level of performance enough to make them unattractive...

Why in the world are they still working on 5 discrete chips?

I could maybe see 3 desktop chips + 1 mobile only chip (with the other 2 mobile chips using the bottom 2 desktop chips like AMD). But 5 chips seems silly.

Regards,
SB
 
I agree with the 5 chip question - I would think that at the very low end, the market for the very lowest end cards is going to start overlapping with integrated graphics platforms more and more, as people upgrade their motherboards / CPUs. I can see some pick-up from people who still want basic graphics on an old motheboard, but that has to be a rapidly shrinking segment. I would think that 3 desktop chips could cover the entire market pretty well.
 
sethk said:
I agree with the 5 chip question - I would think that at the very low end, the market for the very lowest end cards is going to start overlapping with integrated graphics platforms more and more, as people upgrade their motherboards / CPUs. I can see some pick-up from people who still want basic graphics on an old motheboard, but that has to be a rapidly shrinking segment. I would think that 3 desktop chips could cover the entire market pretty well.
Click to expand...
How much faster is the fastest integrated SB SKU compared to a GF119? I honestly don't know, but if the lowest end GK chip is faster by the same ratio, there should still be a lot of interest, I think. Especially since Intel drivers continue to be low quality.

3 chips is definitely too low. ~500mm2 is a necessity for HPC. ~350mm2 is for the mainstream performance segment. That gives you a gap of, say, 80mm2 to 250mm2 to fill with just one die. I don't see how that's doable without giving up at least some major market.
 
jaredpace said:
My guesses are the GK104 part coming out in April-May is this part:
This chart from Dec/early Jan may be correct chart:
9h6xe9.png
Click to expand...
I'm not convinced. I think just about the only thing nvidia promised was much better DP flops/w and that will be difficult to achieve if going back to 1/3 DP/SP ratio on the high-end (I certainly understand why it ended up that way in the chart with assuming similar SMs this time around for all parts).
Also the alu/tmu ratio would be a big jump (1.5 times that of the already quite high GF100/GF110 and twice that of the other GF1xx parts) at the high end while the low end keeps the same ratio as the non GF100/GF110 parts. Not sure if this makes sense (but again of course this is because the SMs are assumed similar).
The projected 3dmark11 scores make no sense at all when comparing GK106 and GK104 as the percentage difference is actually higher than any of the individual key stats (when taking into account clocks). Even disregarding that "minor" rounding error it would mean GK106 is 100% alu-bound in 3dmark11 and it simply makes no sense to build parts which are that unbalanced.
So all told looks about as random guess as anything else I've seen.
 
silent_guy said:
How much faster is the fastest integrated SB SKU compared to a GF119? I honestly don't know, but if the lowest end GK chip is faster by the same ratio, there should still be a lot of interest, I think. Especially since Intel drivers continue to be low quality.
Click to expand...
GF119 is still faster than any SB graphics (not by that much mind you).
I think it is a losing battle though going forward for low-end chips. Saddled by 64bit memory interfaces, they either are at a very big bandwidth disadvantage (with ddr3) or an even bigger power disadvantage when using gddr5 (certainly that's only a problem for mobile not desktops but still).
So GF119 being faster than SB IGP is already amazing enough (with its low ddr3 memory bandwidth and no large cache it can use compared to SB).
It is probably a bigger problem for amd (which want to sell APUs instead) but I'm not terribly convinced those low-end chips still make a lot of sense for nvidia neither.

3 chips is definitely too low. ~500mm2 is a necessity for HPC. ~350mm2 is for the mainstream performance segment. That gives you a gap of, say, 80mm2 to 250mm2 to fill with just one die. I don't see how that's doable without giving up at least some major market.
Click to expand...
Agreed but 4 chips might do. AMD already dedicates more than 80mm² to the graphic part of their APU and intel is probably into similar transistor count for their ivy bridge igp compared to any such 80mm² chip (of course being 22nm it will be smaller) - and that figure certainly doesn't include the LLC.
Of course there are still cpus which don't have IGPs but I would guess for most of them either a higher end card is in order or mostly any old chip will do.
 
mczak said:
I'm not convinced. I think just about the only thing nvidia promised was much better DP flops/w and that will be difficult to achieve if going back to 1/3 DP/SP ratio on the high-end (I certainly understand why it ended up that way in the chart with assuming similar SMs this time around for all parts).
Also the alu/tmu ratio would be a big jump (1.5 times that of the already quite high GF100/GF110 and twice that of the other GF1xx parts) at the high end while the low end keeps the same ratio as the non GF100/GF110 parts. Not sure if this makes sense (but again of course this is because the SMs are assumed similar).
Click to expand...
I don't wanna touch 3dmark scores, but what i heard whispering was close to this table (no hotclocks, chip/mem clocks very close to Tahiti, dp/sp ratio for non-tesla products). Also, very confident tahiti is no problem performance wise. Could be a bluff tho.
 
About the 5 chips being too much: I think markets such as LGA2011 & AM3+ low-end graphic card will always exist. So the lowest end GPU sales will just shrink in volumes, but are here to stay.
 
Silent_Buddha said:
What I don't get is that with the death of the budget market due to integrated graphics approaching their level of performance enough to make them unattractive...

Why in the world are they still working on 5 discrete chips?

I could maybe see 3 desktop chips + 1 mobile only chip (with the other 2 mobile chips using the bottom 2 desktop chips like AMD). But 5 chips seems silly.

Regards,
SB
Click to expand...

If this table is correct—which it probably isn't—there's a factor of 12 between the biggest and smallest chip, considering the number of shaders. If you take the GeForce 200 series, for instance, there was a factor of 15 (16-240 shaders).

The gap between the top chip and the "performance" one has shrunk, and the bottom end has been pulled closer to the rest, so things have changed a little. As long as the smallest chip still beats Ivy Bridge by a comfortable margin, it makes sense.
 
they can make cheap Quadros, which they do already. nice low power, high margin cards for Autocad and the like.
perhaps importantly, not making the small chips would mean you're giving up. not the greatest thing to do.
 
S|A - GK110 tapes out at last

He said that GK110 is basically reticle limited, about 23.5mm on a side. The math says that it is about 550mm^2, with the last two generations coming in at 529/550mm^2 (GF100/GF110 respectively) and 576mm^2 for GT200. [...]One thing the mole said when he looked up with somewhat bleary eyes is that the chip definitely has a 384-bit memory bus.
Click to expand...
 
mczak said:
I'm not convinced. I think just about the only thing nvidia promised was much better DP flops/w and that will be difficult to achieve if going back to 1/3 DP/SP ratio on the high-end (I certainly understand why it ended up that way in the chart with assuming similar SMs this time around for all parts).
Also the alu/tmu ratio would be a big jump (1.5 times that of the already quite high GF100/GF110 and twice that of the other GF1xx parts) at the high end while the low end keeps the same ratio as the non GF100/GF110 parts. Not sure if this makes sense (but again of course this is because the SMs are assumed similar).
The projected 3dmark11 scores make no sense at all when comparing GK106 and GK104 as the percentage difference is actually higher than any of the individual key stats (when taking into account clocks). Even disregarding that "minor" rounding error it would mean GK106 is 100% alu-bound in 3dmark11 and it simply makes no sense to build parts which are that unbalanced.
So all told looks about as random guess as anything else I've seen.
Click to expand...

Since it's the infamous chiphell chart don't bet anything yet, because for some cases it could be closer to reality than many could imagine. As for the 3dmark scores as you can see in a former post above at least one has been circulating in the rumor mill for quite some time now, with the other two being most likely false assumptions since the chart has the GK110 so wrong (starting from the codename) that it's as silly as it can get. Someone from chiphell was merely reading german 3dcenter forum speculations and wrote a chart based on those; absense of any top dog details, it's no wonder that the happy go merry guesswork went as wrong as it did.
 
Ailuros said:
Since it's the infamous chiphell chart don't bet anything yet, because for some cases it could be closer to reality than many could imagine. As for the 3dmark scores as you can see in a former post above at least one has been circulating in the rumor mill for quite some time now, with the other two being most likely false assumptions since the chart has the GK110 so wrong (starting from the codename) that it's as silly as it can get. Someone from chiphell was merely reading german 3dcenter forum speculations and wrote a chart based on those; absense of any top dog details, it's no wonder that the happy go merry guesswork went as wrong as it did.
Click to expand...
Hmm ok leaving out "GK100" and the 3dmark11 scores (from which only presumably the X3000 for GK104 could have any real basis) it sounds more plausible. And seeing GK104 is apparently (that die size story sounds believable to me) pretty much the same size as Tahiti (not some small miracle) I could believe it would be similar in performance too.
If the SA story is correct though it would mean nvidia screwed up the top chip twice in a row for essentially the same reason, except this time they just ditched it instead. Maybe they rather figured risks are too high to go for such a large chip first and just postponed it, in any case the rumors of GK110 only launching Q3 seem to get more widespread...
 
mczak said:
If the SA story is correct though it would mean nvidia screwed up the top chip twice in a row for essentially the same reason, except this time they just ditched it instead. Maybe they rather figured risks are too high to go for such a large chip first and just postponed it, in any case the rumors of GK110 only launching Q3 seem to get more widespread...
Click to expand...

Was the interconnect "broken" on GF110? Think about it and draw your own conclusions.
 
Hate on charlie all you want, he's got better info than a lot of whats posted here usually and he's got info before anyone else usually, too. I watched him tell someone (senior in AMD) at AMD's Tahiti tech day what the next three generations codenames were, and the response 'Where do you get this info?!?!?!!'.

As for the 'backpedalling' you could say that, or you could say 'refining' as he gets better info digging into his stack of moles. He doesn't present it in a dry, passionless/emotionless manner but that's because he's not a peer-reviewed scientific paper.

I find it very interesting that GK110 is allegedly being brought forward to this year to fill in where GK100 won't be. AMD is 'making hay while the sun shines' with Tahiti.
 
That's a little ambiguous in text.
There's :
"Egad! (monocle flies out) Where do you get this info?"
and
"Pshaw. Where do you get this info?"

If the former is how the AMD employee reacted, remind me to schedule a poker game with them.
 
You must log in or register to reply here.

Similar threads

A
RDNA3 Efficiency [Spinoff from RDNA4]
3 4 5
Replies
85
Views
9K
Bondrewd
Bondrewd
D
The Silent Hand of Driver "Optimization" (spinoff)
2
Replies
31
Views
3K
DavidGraham
D
F
RDNA5 Speculation
Replies
2
Views
1K
Erinyes
E
Shortbread
Nvidia is Now The Second-Largest Company in The US Market
Replies
1
Views
3K
Shortbread
Shortbread
T
GPU Stability testing for Nvidia cards - PowerSweep
Replies
0
Views
453
T2098
T
Back
Top