NVIDIA Maxwell Speculation Thread

http://www.overclock.net/t/1464898/...enchmarks-surface-in-the-uk/200#post_21799020

GTX 750 Ti can overclock to 1.2GHz without external power connector.

Standard Disclaimer: Not all chips can overclock that much. Your milage may vary.

 

They don't add anything to flops count so it doesn't make much sense to count them. Well actually with G80 they did contribute theoretically though just a tiny bit (they could do mul at IIRC 1/4 rate that of the ordinary "cores", much controversy about the "missing mul"). I think with Fermi/Kepler though this was no longer possible at all since it was impossible to dispatch enough instructions simultaneously to SFUs and ordinary ALU lanes (and I don't know or can't remember if they actually still could do a mul). I suspect that's still the same with Maxwell, though I may misinterpret the shiny diagrams.

There aren't any SFUs since those functions are all integrated into ordinary ALUs. Well actually I don't know how integrated they really are but simultaneous dispatch is definitely not possible.

 

Counting SFUs where: In GFLOPS or reported number of ALUs?

The answer is neither. There are 64 ALUs per GCN CU. (To be precise, there are 16 ALUs for each of the 4 SIMDs within a CU.) Each ALU is float mad capable and have the same float mad rate. It's quite easy to compute the peak GFLOPS and you can even make tests to hit that rate.

 

Nice point. It's too bad, as I'm more interested in hardware acceleration of 2-pass encoding than real-time, but low-power, real-time encoders, particularly if they start supporting vp9/H.265 or any of the 4k formats, would be a nice add for data acquisition use-cases as well.

 

Wolfram alpha gives 12.1% increase for (0.9/0.85)^2.

What am I missing here?

 

hmmm...
galaxy-geforce-gtx-750-ti-exposed-can-overclocked-1-35-ghz
...impressive OC

 

http://www.youtube.com/watch?v=9SnXDPe2gSY

Reference card noise test.

 

Memory needs to overclock well for that 1.35 ghz to matter. I feel like this chip (at current memory speeds) is going to be bandwidth limited much the same way gtx 680 was.

EDIT: 7.0 ghz vram I see. That will do. Haha.

 

My only explanation is that I must have typed 0.95 instead of 0.90…

 

When do the reviews embargo lift?. Maxwell seems an exciting architecture.

 

Why is the marketing material mainly comparing the card to the GTX 550 Ti, which is a 40nm chip with half the transistors?
Sounds like they're avoiding the comparisons to the GTX 650 Ti...

 

First independent reviews should be only hours away; then we'll be quite a bit wiser.

 

There are new "PR benches" out at videocardz comparing the card to r7 260 and 260x as well as GTX 480. Anyway we should get proper reviews in a couples of hours
I'm actually quite excited those cards could do great in the htpc I'm dreaming to build (/money ).

 

Because marketing has found that most gamers don't upgrade every cycle but every two cycles so that comparison is directed to them.

Tin foil hat comment.

 

Link please as I do not see said "PR benches" on the front page.

I am looking to replace my GTS 450 in my HTPC with one of these cards.

 

I think they are here

 

It doesn't really matter what you call them since the results most likely come from NV.

 

the 750 Ti's main competitor should be the R7 265 (not the 260 as in NV's PR) as this new review suggests:

http://www.techpowerup.com/reviews/Sapphire/R7_265_Dual-X/24.html

 

Thanks, looks like they updated the article recently with the AMD comparisons.

 

The R7 265 didn't exist until 5 days ago and Nvidia's slides were made some time ago so you can take off your tin foil hat now.