NVIDIA Kepler speculation thread

[fellix]

Another mirror for the slides: http://www.bouweenpc.nl/nieuws/hardware/review-nvidia-geforce-gtx-680-uitgelekt

Nevermind the nasty watermarks. >_>

 

[SimBy]

I wonder what happens at a normal 4xAA and higher res. It might even be behind HD7970. At least in
a few games. Might even be further ahead who knows.

 

[silent_guy]

No word on real power consumption and overclockability (it is even possible?).

No word necessary: it's impossible to engineer such thing to have the mass production version skirt the edge of what's possible. You •have• to add a guard band to avoid excessive returns. It's probably different with the x990 and x90 cards because those are low volume affairs anyway, but you cannot afford to be inundated with returns.

 

[oramay]

Thank you, sir, but: "Imgur is over capacity!"... d'oh!

Seriously? In any case, I'm going to bed now. Good night.

 

[tviceman]

From what I could see it dynamically overclocks to around 1100 MHz in some games and benchmarks.
Base clock is 1006 Mhz. Package size is 18.55 x 17.18 mm. 3.54B trannies.

All benchmarks are at 1920x1080 8xAA and biggest advantage I've seen was around 20%.

HD 7970 at 1.2GHz would be on par or slightly ahead.

17.18 x 18.55 comes out to 318 mm^2. So is that with the heat spreader on? The slide on that site still says it being 294 mm^2.

 

[AnarchX]

Manufacturer die-sizes are always a bit below, that the ones you can measure on a package.

 

[Mize]

I wonder what happens at a normal 4xAA and higher res. It might even be behind HD7970. At least in
a few games. Might even be further ahead who knows.

I'd say you covered the bases there.

 

[Creig]

It loads ok for me so far. 680 is cooler and perform better than 7970.

That slide shows the 580 as having the exact same load temp as a 7970. All of the reviews I've seen put the 7970 anywhere from 4°C to 10°C cooler under load. Seems fishy.

 

[Love_In_Rio]

17.18 x 18.55 comes out to 318 mm^2. So is that with the heat spreader on? The slide on that site still says it being 294 mm^2.

That measurement includes the encapsulator, the chip itself will be smaller.

 

[Man from Atlantis]

Txaa, 128 tmus, 256 sfus, adaptive vsync etc there is also a dieshot

 

[fellix]

I wonder why they left the LDS/L1 combo size inchanged?

 

[Love_In_Rio]

With only 0.5 million transistors more than GTX 580 they get 40-50% more performance. So, how?. Have they ditched DP logic?.

 

[Arty]

No mention of cache sizes?

Now I wonder why AMD's boards are longer when Nvidia has faster memory and yet has shorter boards.

 

[Love_In_Rio]

No mention of cache sizes?

Now I wonder why AMD's boards are longer when Nvidia has faster memory and yet has shorter boards.

Because Nvidia has better cooling solution?.

 

[Arty]

Because Nvidia has better cooling solution?.

Wasnt it due to requiring longer traces on the board?

 

[Man from Atlantis]

SMX reminds me old TPC but CCs were seperated 3 parts there.. Fermi is 1, Kepler 2 and Maxwell will be 3 again through the Evolution to Einstein i hope

 

[fellix]

Apparently, Kepler architecture is simply exploiting the opportunities in the 28nm process, applying the same old Fermi "magic" with some new power managing features. Maxwell should start another new cycle of fresh architectural design all over again.

 

[Erinyes]

With only 0.5 million transistors more than GTX 580 they get 40-50% more performance. So, how?. Have they ditched DP logic?.

I assume you meant Billion. But anyway the right comparison would be to GF114. It has 80% more transistors than GF114 (1.95 Billion), and about 70-80% more performance.

Here are the slides and photos for people who have problems loading.
http://imgur.com/a/aQmuA

Thanks!

On one of the slides they mention "8 Geometry Units". But I see only 4 GPC's?

 

[tunafish]

Here are the slides and photos for people who have problems loading.
http://imgur.com/a/aQmuA

Thank you.

 

[tunafish]

With only 0.5 million transistors more than GTX 580 they get 40-50% more performance. So, how?. Have they ditched DP logic?.

On paper, the process is supposed to be 40% better (faster transistors). It's a HKMG process vs bulk.

The big question is not why is it so good, it's why Tahiti sucks so much.