NVIDIA Kepler speculation thread

AMD is beat on performance/watt and performance/mm2. Finally!

 

Higher AA should favor the card with higher ROP and memory throughput theoretically. Higher resolution's usually good for architectures with massive fillrates, the memory size and architecture to support it and is not as dependant on a very efficient front-end, as triangle-sizes go up.

 

And in the end both will also will beat the consumer in performance/price.

 

I think it's about on par with Pitcarin. Which is impressive.

 

There are 3 numbers possible:
This one: includes active die, includes scribe around active die, includes some amount of glue around the scribe
Die size as used to calculate dies per wafer: glue not included.
Die size as used to calculate yield: scribe not included. This is usually the number used most internally.

Not unreasonable to go from ~300mm2 to ~320mm2 to go from one to the other.

 

Actually, there is a difference between 4x and 8x MSAA. As AMD ROPs can't handle 8xMSAA at the speed NV ROPs are capable of, the ROP load basically doubles at 8x for AMD cards (but not for nV). The memory bandwidth plays a bigger role at 4x in case of AMD (because it is less limited by the ROP throughput itself), at 8x it doesn't help that much. So at 4x the bandwidth can have a potentially larger effect than at 8x (while in nV's case a higher bandwidth should pay off more at 8x).

 

Tahiti does not suck. Why do people always say that it sucks? I mean, ok, it could be cheaper, but that's really all that kept me from buying one in the last two months.

 

This is AMD's most compute-focused GPU yet. In order to compete with Nvidia in this regard, they had to take a hit somehow. Nvidia has been doing compute-focused architectures for 3 generations now, whereas Cayman was AMD's first, but not fully-focused, attempt. Nvidia has more experience in this regard.

 

The Polymorph 2.0 Engines are the "Geometry Units".

And all the game benchmarks given are a 8x AA. Can we infer 4xAA and noAA performance from that?

 

Actually, what I'm seeing is different. Comparing the latest High-End cards that both have 384 bit of memory interface and where Nvidia has the advantage of 50% more ROPs, AMD looses less performance when going from 4x MSAA to 8x MSAA.

 

can anyone capture the die shot without watermark

ok i found it

Spoiler

 

I believe the "Geometry Units" are called "PolyMorph Engine 2.0" and they are 8 of them one per SMX.

http://imgur.com/a/aQmuA#EFjJN

http://imgur.com/a/aQmuA#EXqKU

 

The L1 cache size is 512 KB. 64KB per SMX. 8 SMX's.

The Uniform, Texture and Instruction Cache sizes are unknown.

http://imgur.com/a/aQmuA#EXqKU

Wouldn't the wider memory bus size and more memory chips result in longer boards?

 

Cleaned for thou geekish pleasure :



The command processor in the middle is strangely smeared. Looks like NV doesn't want anyone to peek in the gigathread factory.

 

L2 is 768

 

Link please?

 

More like a sloppy graph editing.

 

I remember a shot of this naked kepler die that had the individual sections highlighted with labels. Can someone provide a link?