NVIDIA Kepler speculation thread

[Alexko]

Err... multiply by 3? Having 3x ALUs doesn't increase performance by 3x, especially when the Fermi ALUs are hotclocked and these aren't, not to mention other possible differences

I think your sarcasmometer needs a tune-up.

 

[ECH]

from 3dc

550mm2 on 28nm
I know this can be all taken into speculation. But, really
I have to wonder, at the very least, if a dual GK104 would provide higher results.

 

[DarthShader]

Kyle says the leaked THG slides are the real thing:

http://hardforum.com/showpost.php?p=1038514895&postcount=54

Nice job, THG. :S

 

[Man from Atlantis]

550mm2 on 28nm
I know this can be all taken into speculation. But, really
I have to wonder, at the very least, if a dual GK104 would provide higher results.

2x195W vs < 300W card, it's obvious..

 

[trinibwoy]

This is a non-argument. Same applies to AMD's stuff.

Huh? The HD4890 ran at 850Mhz. AMD has been up in that range for a long time.

 

[lanek]

Sandra gpgpu bench against tahiti AND pitcain:

http://www.theinquirer.net/inquirer...hed-amds-mid-range-radeon-hd-7870-gpu-compute
Wonder if the clock is really 700 or 1000 here..

AMD gpus work a lot better in this test ...maybe due to OpenCL.

Here's 2x 7970 stock not comparable with one ( and the score of one was allready in their test ), but this is just for show the 580 and 5870 scores. ( green is the 580, blue is the 5870)

 

[A1xLLcqAgt0qc2RyMz0y]

from 3dc

Some math:

GK104 has 8 SMX's, 1536 cores
GK110 will have 12 SMX's 2304 cores

That equates to the 150% gaming performance

So if the GK110 only has 50% mores cores why does the GK110:

Die size increase by 87%
Transistors increase by 69%

 

[fellix]

AMD gpus work a lot better in this test ...maybe due to OpenCL.


http://imageshack.us/photo/my-images/835/sandragpgpu.png/

NV is artificially throttling DP ALU throughput in GeForce SKUs at 1/4 rate.

 

[CarstenS]

from 3dc
http://i42.tinypic.com/akh9w3.jpg

Not all of those numbers are real.

 

[Alexko]

Some math:

GK104 has 8 SMX's, 1536 cores
GK110 will have 12 SMX's 2304 cores

That equates to the 150% gaming performance

So if the GK110 only has 50% mores cores why does the GK110:

Die size increase by 87%
Transistors increase by 69%

I wouldn't know about the validity of those figures, but I'd expect GK110 to have:

• more registers per SP,
• possibly narrower SMXs, just like GF100 vs GF104,
• more L1 cache per SMX, or at least per SP,
• more L2 cache,
• a wider memory bus,
• ECC everywhere,
• 1/2 DP support,
• perhaps various other features that GK104 lacks.

In fact, if you look at Tahiti, it's bigger than what you'd expect from looking at Pitcairn and its SP count, and a lot bigger if you also consider that it has the same amount of ROPs, a very similar front-end, and a memory bus that is "only" 50% larger, just like the L2 cache.

 

[fellix]

I wouldn't know about the validity of those figures, but I'd expect GK110 to have:

• more registers per SP,
• possibly narrower SMXs, just like GF100 vs GF104,
• more L1 cache per SMX, or at least per SP,
• more L2 cache,
• a wider memory bus,
• ECC everywhere,
• 1/2 DP support,
• perhaps various other features that GK104 lacks.

Altering fundamental aspects in the micro-architecture (at SM level) within a generation is unlikely.
On the other hand, with such fat multiprocessors, I simply can't imagine what a 16 SMX Kepler part will look like. It would have very weird packaging - something like the GF104 die, but far larger.

 

[Erinyes]

See, you can't assume that they have no big chip unless you assume that GK104 was always meant to occupy the flagship role. Now ask yourself how in the world could nVidia have assumed that a < 300mm^2 chip could beat AMD's best when they've been so far behind the perf/mm curve for several generations in a row? If we want to assume stuff then the assumptions should at least make sense!

It's a rather pointless argument anyhow, I think the facts (will) speak for themselves.

You are assuming that Nvidia assumed? NV likely could not have known what AMD was planning, nor what performance AMD would end up with (especially given the timeframes involved in the design of these chips). They did their own thing and as it turns out did a very good job. So come launch time they could launch their intended mid range card against AMD's high end.

A more interesting question is "why didn't NVidia do it this way from day 1, i.e. G80".

They kind of did the same thing with G92. It launched before GT200 on the same process, and arguably its performance would have been closer to GT200 had it been designed with a GDDR5 memory controller.

550mm2 on 28nm
I know this can be all taken into speculation. But, really
I have to wonder, at the very least, if a dual GK104 would provide higher results.

Whats so surprising? They launched Fermi first on 40nm and it was 530 mm2. What makes you think they wouldnt go for a large die this time around? Their strategy is more conservative this time around as they decided to wait till the process was more mature before going for a big die.

Sure dual GK104 wll provide higher performance, but GK110 has dual role as a HPC card, which GK104 cannot provide.

Some math:

GK104 has 8 SMX's, 1536 cores
GK110 will have 12 SMX's 2304 cores

That equates to the 150% gaming performance

So if the GK110 only has 50% mores cores why does the GK110:

Die size increase by 87%
Transistors increase by 69%

Higher DP rate, other HPC features, Larger caches, ECC, Wider memory bus are some of the things i can think of off the top of my head.

Edit: Alexko beat me to it, and yes the Pitcairn v/s Tahiti analogy is what i was thinking of as well for GK104 v/s GK110

 

[DarthShader]

Altering fundamental aspects in the micro-architecture (at SM level) within a generation is unlikely.

Indeed, but the GK110 is kinda next generation, isn't it? The caches remained comparable to a GTX580, while theoretical FLOPS doubled...

 

[Broken Hope]

Have these been posted already?

http://www.youtube.com/watch?v=68sRjUilQHw
http://www.youtube.com/watch?v=kzI4an_cllA

 

[Man from Atlantis]

GeForce GTX 680 Launch Video

 

[trinibwoy]

NV likely could not have known what AMD was planning, nor what performance AMD would end up with (especially given the timeframes involved in the design of these chips). They did their own thing and as it turns out did a very good job. So come launch time they could launch their intended mid range card against AMD's high end.

That's exactly what I was implying

 

[Rangers]

Not at all. It doesn't have to "soundly beat" it for us to see that the balance has shifted from last generation.

It has indeed, AMD is now closer than before to performance leadership.

 

[Broken Hope]

What are the chances of AMD lowering prices now? or are they going to use the Eyefinity/High resolution performance differences to justify being the same price?

 

[CarstenS]

A more interesting question is "why didn't NVidia do it this way from day 1, i.e. G80".

Until Fermi, Nvidias main battle was against area because they could afford the big die strategy due to their near-monopoly in the professional market. Quadro cross-financed Geforce. Now their main enemy is power and newer process technology has been proven to be unpredictable at best, unreliable at worst in the first couple of month after it's inception.

Obviously not being confirmed by Nvidia, I am under the impression, that they did not hope to best Tahiti with GK104 in the first place.

 

[Erinyes]

Not all of those numbers are real.

Only the 1636 CUDA cores for GK104 or anything else?

What are the chances of AMD lowering prices now? or are they going to use the Eyefinity/High resolution performance differences to justify being the same price?

If GTX 680 is at $499 as rumoured..AMD will have to drop 7970 to $449 at the very least. Lets hope so!