NVIDIA Kepler speculation thread

pjbliverpool said:
I hope it's true but if the 760Ti is basically just a rebranded 670 then Titan will be around 50% faster than it. And I'm not sure we'll get a 50% performance improvement between the 760Ti and 860Ti. It's possible but I wouldn't put money on it.
Click to expand...

I have to clarify that I didn't mean any particular model number or pricing level. Basically I meant that their second biggest Maxwell chip in it's top configuration should/could be faster than Titan similarly to GK104 vs GF110. Titan is such a beast though that I'm far from sure about that. It has a bigger lead over GK104 than what GF110 had over GF114, even with only 14/15 of the chip activated.
 
Kaotik said:
They no doubt could, but they won't, they don't see the point in monolithic chips. Next gen will be most likely 20nm, they shouldn't have any issues making chip faster than GK110 with it without going monolithic chips.
Click to expand...

No need for a larger monolithic die. Imagine if they beefed up Pitcairn instead of Tahiti. For the consumer space that would be good enough, DP and whatnot aren't exactly critical in that space. Take the Pitcairn arch. and bump it up to Tahiti size and I'd imagine that could potentially be faster than Titan in gaming workloads.

Will it happen? I have no clue. As I have no clue how difficult that would be to accomplish. :) Hell, bump up Pitcairn to GK104 size and it might be able to match Titan in performance. :) Considering Keplar has nothing on Pitcairn when it comes to performance/mm^2

Regards,
SB
 
Clock for clock the 7970 Ghz scales quite well over the 7870. Pitcairn doesn't have any special magic over it.
Pitcairn has 58.1% of the area and 62.5% of the CUs compared to Tahiti with only marginally better performance than what the CU difference would imply, and the edge shrinks as the graphical settings intensify.

7970Ghz is fine as it is, it just wasn't nearly at its current level when it launched with regards to clocks and driver performance.
 
Dr Evil said:
Clock for clock the 7970 Ghz scales quite well over the 7870. Pitcairn doesn't have any special magic over it.
Pitcairn has 58.1% of the area and 62.5% of the CUs compared to Tahiti with only marginally better performance than what the CU difference would imply, and the edge shrinks as the graphical settings intensify.

7970Ghz is fine as it is, it just wasn't nearly at its current level when it launched with regards to clocks and driver performance.
Click to expand...

It's obvious that Tahiti-sized Pitcairn would be running circles (exaggerating a bit obviously) around Tahiti, and based on the scaling of the architectures on the chips which don't have similar focus on GPGPU performance as Tahiti, it should beat GK104 silly mm^2 to mm^2 too
 
If it's so obvious, why don't you bring out some numbers to back that up? More like wishful thinking imo.

GTx 680 is 38% bigger than Pitcairn and it's very close to being exactly that much faster.

Cape verde is 10CU/123mm2, Pitcairn is 20CU/212mm2 and Tahiti is 32CU/365mm2. If you just divide the area with CUs, you get 10.6mm2 per CU for Pitcairn and 11.4mm2 for Tahiti. 32CUs x 10.6 would be 339mm2 and you'd imagine that Tahiti's memory controller eats up a bit more space too. There is not much room left for a miracle. Maybe they could put one or two more CUs in there.
 
I am not sure, if AMD will be able to [edit: sufficiently] feed 20 CUs or more off of one Setup-Engine.
 
Dr Evil: Are you sure your numbers are correct? Maybe I'm missing something, but I got about 5,47 mm² per CU for Tahiti and 4,40 mm² for Pitcairn.

So, Tahiti (365 mm²) is large enough to fit ~40 Pitcairn-sized CUs (even with its large 384bit memory interface). Maybe 36 CUs with 3 setup engines would be slightly more balanced configuration. Anyway, both of them would have better gaming performance per mm² ratio.
 
Well I'm no math wizard by any stretch of the imagination, but to me it looks my numbers are correct and yours aren't :)

How did you calculate those?
 
I used the Tahiti die-shot to measure one block of 16 CUs and I divided the result by 16 afterwards. Simple. Your result cannot be correct - whole Pitcairn including ROPs, memory controller and interface (etc.) measures 212 mm², so one CU cannot be 10.6 mm² large.
 
My calculation wasn't supposed to show what the size of the CUs are, but that the relative "fat" Tahiti die has compared to Pitcairn is quite small.

Just another way to say what I said a couple posts back "Pitcairn has 58.1% of the area and 62.5% of the CUs compared to Tahiti"
 
Tahiti's main problem is not CU-density, it's power-efficiency, or rather its lack thereof.

Pitcairn is much more efficient, and Bonaire even more so.
 
Dr Evil said:
If it's so obvious, why don't you bring out some numbers to back that up? More like wishful thinking imo.

GTx 680 is 38% bigger than Pitcairn and it's very close to being exactly that much faster.

Cape verde is 10CU/123mm2, Pitcairn is 20CU/212mm2 and Tahiti is 32CU/365mm2. If you just divide the area with CUs, you get 10.6mm2 per CU for Pitcairn and 11.4mm2 for Tahiti. 32CUs x 10.6 would be 339mm2 and you'd imagine that Tahiti's memory controller eats up a bit more space too. There is not much room left for a miracle. Maybe they could put one or two more CUs in there.
Click to expand...
The theoretical area possible per CU is really an odd number, I'd rather look at just the whole chip, and more CUs is NOT the answer, it would do nothing at all for performance (in graphics).
And there it's obvious Tahiti is just bad compared to its siblings, Cape Verde -> Pitcairn gets you ~80% more performance for roughly ~80% larger chip (slightly better scaling in fact). Pitcairn -> Tahiti gets you for another 70% area only ~40% more performance. And as was mentioned power efficiency decreases alongside the perf/area decrease. A "triple-CapeVerde" would be only about ~300mm² and be just as fast (and that's even taking into account that scaling might be slightly worse than CV->Pitcairn).
 
One thing is for sure- given NVIDIA's obvious ability to design better architectures, and if AMD doesn't fix the failures with Tahiti, then Maxwell's version of Titan will be even more painful for AMD- either the performance gap between the top solutions will be bigger, or NV will achieve same performance gaps with smaller chips and will earn tons of more money
 
UniversalTruth said:
One thing is for sure- given NVIDIA's obvious ability to design better architectures
Click to expand...
Really?

I see different, I don't see better. There are legitimate arguments to indicate that the tradeoffs made for Keplar are a blind alley and they will need to bias back for future gen.
 
UniversalTruth said:
One thing is for sure- given NVIDIA's obvious ability to design better architectures, ...
Click to expand...
I'm wrecking my brain trying to see the 'obvious' in your statement.

The GCN and Kepler architectures have their strengths and weaknesses, but, compensating for DP capability, they're pretty similar in performance.
7970 has some minor issues in the power consumption department, but optimal power is often the first thing to fly out of the window when schedule is at risk. (Especially when the part isn't scheduled to have a parallel life in a laptop?) If there's another 28nm iteration, I expect this to be one of the things that will be fixed.
 
UniversalTruth said:
One thing is for sure- given NVIDIA's obvious ability to design better architectures, and if AMD doesn't fix the failures with Tahiti, then Maxwell's version of Titan will be even more painful for AMD- either the performance gap between the top solutions will be bigger, or NV will achieve same performance gaps with smaller chips and will earn tons of more money
Click to expand...

How is making notably bigger "gaming only" chip "obvious ability of designing better architectures"?
 
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
449
T2098
T
Back
Top