Nvidia BigK GK110 Kepler Speculation Thread

NVIDIA pays little attention to OpenCL, there's no reason Titan should change that.

CUDA is the target.

 

Really expensive!
No backplate
Voltage control very limited
Power limit can only be adjusted by +6%
Boost 2.0 adds more complexity to overclocking

http://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_Titan/35.html

 

How about reading the entire page before you post:

AnandTech also notes that it is an OpenCL bug in the current driver. That's why Anand isn't testing with OpenCL.

 

Single GPU, no multi-GPU issues
Very fast
Quiet during gaming
Low power consumption
Good overclocking potential
Sexy high-quality design
Extremely quiet in idle
Boost clock 2.0 adds new overclocking features
6 GB memory
Support for voltage control
Up to four active outputs
Native full-size HDMI & DisplayPort
Adds support for Display Overclocking
Support for CUDA and PhysX

 

Can we debunk this myth ? All Kepler family lacks of good performance in most FP32 GPGPU applications.

Just try to run Blender Cycles or Octane (both CUDA-only off-line renderers) on a GTX 580 and than on a GTX 680.

 

OpenCL and CUDA are very very similar and they perform similar in almost all scenarios.

I just hope for Nvidia that this is gonna be true..

 

It's a joke.
All of these is true for almost every card out there on the market.

 

I guess there are driver issues in a few games still but I think it goes both ways. I notice AMD still lags badly in Shogun II for some reason. But yeah ~30% seems about right, or at least it'll be that once the drivers are sorted.

Very concerned about this boost shenanigans though. Cold cards adding 10% or so performance is just not on, because they won't be cold while gaming. I really believe this needs looking at HARD by the tech press.

 

Look at our review then. Link's some postings further up.

BTW - wrt to GPGPU. I haven't seen 2 TFLOPS of DP-Performance out of a single processor yet - do you?
http://www.pcgameshardware.de/Gefor...orce-GTX-Titan-mit-ueber-2-TFLOPS-DP-1056533/


Edit: Obviously, I need to point out that this only possible with overclocking. There, I said it.

 

The problem with Kepler is that it doesn't have a lot of register space and cache vs. its computing resources, so it tends to choke on somewhat complex workloads.

CUDA lets you tap into the TMU memory to alleviate this problem, but as far as I'm aware, OpenCL doesn't. Even so, it requires a bit of extra effort, and typical CUDA applications don't necessarily do it.

 

Possible? The specs say 1,3-1,5 TFLOPS DP

 

The aggressive power and temperature throttling is a surprise, and it makes sense, the default is not noisy and there's not too much heat to get rid of. You can install it and forget about it if you wish.

Value for the price is dubious of course (give me a vanilla GTX 660 and I'll be happy) unless you're a PhD who badly needs such a new toy, 6GB Cuda with 1.5 teraflops DP.

Deal with it! It's hardware you can't afford, so pretend it's not there and you will be fine.

I may compare it to the iphone with max flash storage, at 900 euros (included taxes). Rofl!
that computer phone is terribly overpriced, isn't it.

 

Did you hear of something called Overclocking?

 

http://hwbot.org/submission/2359033...me_preset_geforce_gtx_titan_4352.8_dx11_marks

1,685MHz on LN2, not bad at all.

 

I agree regarding the caches, I dont regarding registers

http://www.realworldtech.com/kepler-brief/

You can use TMU only under special conditions.

 

Anyway, any place to get this OCL benchmark ? cant find it or only bad / removed download link. ( even by Inpei )

 

Ach so

I skipped at the spot, this time I saw it all, very nice

 

Why don't you agree?



Register file size per SP flop per SM(X) or CU:

GF100: 128/64 = 2
GF104: 128/96 = 1.33
GK104: 256/384 = 0.67

GCN: 256/128 = 2

 

Boost 2.0 probably is somewhat bugged now as it works based off temperature and less power consumption as GTX 680 boost 1.0 was. As you said wait till this is sorted via drivers. There maybe no deliberate shenanigans as you seem to be implying....

 

If it's based on temperature, lowering clocks when the card gets hot (and increasing them when it's cold) doesn't seem like a bug to me, it seems like feature doing exactly what it's supposed to.