Kepler test build [Obsolete]

[x3studio]

@t_3 The real raw performance (like hardware sheduler) is missing in "new" design! Also FP64 is poor (some use this for other than graphics compute tasks). It is more than good for gamers, though. I'm almost sure that GTX685 (780?) will be on same compute power level as GTX 590 (but more efficient per Watt, as only 1GPU will deliver comparable results as 2 GPU).

[Timmaigh]

That is one big FAIL of NVidia!

If we consider that Radeon 7990 will have 7.6 TFLOPS Single Precision compute power and 1.8 TFLOPS Double Precision compute power - then NV is seriously in deep sh%%t.

only that the 7990 is a dual gpu solution...

if we consider that the upcoming gk110 will have 2tflops dp power from a single gpu, and the gk104 already generates a little over 3tflops raw sp performance out of a single chip, i don't think that nvida will need to worry seriously or deeply esp. because for large scale gpgpu computing people usually count flops per watt, thus the gk110 might be a clear winner in this area.

the more interesting question is, how long it will take until cuda/compilers/engines will take better advantage of the raw kepler performance, and what is in the end possible with the current gk104 design. afaik the gk110 will not only have more shaders, but will again have a different design (and notably higher prices), but most probably better suited for gpgpu... and the question is, if nvidia cares for "small business" gpu rendering needs

Lets hope so, otherwise i will whine on Nvidia forums to no end....

[gabrielefx]

Yesterday I purchased 4 new GTX580 3GB

more ram, more power, more heat....

[be_fast]

Yesterday I purchased 4 new GTX580 3GB

more ram, more power, more heat....

Hey man, how are you going to use them? Super-micro mother board? Please share you thoughts. As I see you already have 4 cards and you`ll get 8 total. I hope you will not simply built another PC cos it`s not interesting at all )

[Timmaigh]

Yesterday I purchased 4 new GTX580 3GB

more ram, more power, more heat....

Congrats, that is some serious power you have at your disposal now...

On other hand, just yesterday i found out that RAM are DIRTY CHEAP these days!!! 2 years ago 12GB of Kingston 1333 CL7 DDR3 RAM cost me almost 400 EUROs...yesterday, as i could not resist , 24 GBs of Kingston 1600 CL9 cost me 150 EUROs... seriously, that too good offer, i dont think its going to be any cheaper. So i got it, despite not needing it for Octane right now. Maybe in future though, with some 4GB Kepler...

BTW is there a continual proportion between the amount of RAM needed for voxelising and amount of videoRAM needed for scene/ size of the OBJ file? Or can 2 same sized OBJ require vastly different amounts of system RAM to voxelise? I am still yet to buy the Max Plugin, how does voxelising work there anyway? I suppose you need more RAM when using it, than you need with standalone Octane, as you can shut MAX down to make room within system RAM... with the plugin you basically need to accomodate both Octane and Max at the same time, or is the RAM usage completely different?

Sorry for going bit off-topic, but there is not much to add to whole situation for now anyway...

[roeland]

OK, so the current conclusion is that it works, but very slowly. There are some important differences between the Fermi and Kepler architectures, so I expect that once we have a Kepler card and can do the tuning and optimizations needed, we can improve this performance. It is too early to know how good the performance eventually will be.

--
Roeland

[pixelrush]

Perhaps @radiance can contact Nvidia for us and ask for an official explanation of the poor performance.
Its hard to imagine the lame results are due to less than optimal software.
I'm afraid if future Geforce cards are all going to be crippled for cuda then it won't be good for ongoing sales of Octane.
@radiance might have sold at the right time after all.

[glimpse]

Interesting. but things do turn to the side that some of us expected - why would Nvidia give away for free that much of power? =)

Even if they will decide to cut that gains for GTXs and leave the best cuda functionality only for pro series Octane's sales might be hurt, but..untill then, probably not so long from now we are going to see Octane getting out of beta stage.

Then it could actually make sence to buy extra large Quadro paired with tesla cards and render out unthinkable scenes with it =) Yeah It will not be very cheep, but cheaper than building a small render farm for cpu rendering. This path is still an option for some. =) after all, somewhere inside I still hope that GTXs potencial will be unleashed.

[GR1F1TH]

I think it's way too early to tell still. I really wish Refractive had gotten their hands on a 680 before they sold out. The shutdown at TSMC is going to cause a long wait to start seeing a refresh of supply for the 680 and then another long wait for Refractive to optimize. If they get one now them they could have the kinks worked out by the time we can actually buy one.

[mainframefx]

Just look at this and it's totally clear why this card is so slow:
(sorry for the german description)


The SIMD units are an important factor and Kepler got only 8 of them. The HD7970 has 32 thats why it's so much faster in Luxmark.