boxleitnerb
Regular
What does a 580 get there?
Nothing, it's limited to 2 displays.
Man from Atlantis
Veteran
i could try if i had one, my 460 failed hard thanks to 1GB vram 1-3 FPS mostly :smile:
anyone have 1080p screen try to create custom res to run 5760x1080 flawlessly with 1 card and 1 screen
boxleitnerb
Regular
Well, then SLI and divided by 2 should be a good start.
Never mind:
Heaven 2.5 Basic, GTX580 3GB SLI, 5760x1080:
http://www.youtube.com/watch?v=HXyUQb1V0kc
Avg fps: 37.4
Last edited by a moderator:
So the obvious question is, how does it compare to 3x1080p on HD7970 AND was that 29 FPS measured on the central or surround displays
According to this site http://www.overclockersclub.com/reviews/powercolor_lcs_hd7970/12.htm
HD7970 gets 22 fps.
GTX590 gets 27.
Tnx! Impressive if true and without tricks.
Man from Atlantis
Veteran
Heaven 3.0 is released recently, not sure what version the guy has been using..
boxleitnerb
Regular
The 7970 seems a bit slow there?
That would make GK104 60% faster than a single 580.
That would make GK104 60% faster than a single 580.
22fps on an hd7970 running a triple monitor setup with extreme tessellation seems slow?
The resolution only increases pixel and not tessellation load, playing into HD7970's hands.
OT of Heaven 3.0: Results I've seen in the lab indicate that the 3.0 version seems to be quite a bit "faster", i.e. achieving higher fps - at least on some sort of cards.
OT of Heaven 3.0: Results I've seen in the lab indicate that the 3.0 version seems to be quite a bit "faster", i.e. achieving higher fps - at least on some sort of cards.
boxleitnerb
Regular
I thought very high resolutions are a strongsuit of AMD cards in general. But then again, 22fps would make it roughly 30% faster than a 580, so that would be normal I guess.
Well, somebody once claimed that it is fundamentally impossible to have 4x distributed geometry processing without having massive latency and gargantuan power consumption. That same someone inexplicably continued to prove this by pointing at, wait for it: a cell characterization problem that was fixed with a metal patch. (WTF?) Then he murmured something about crossbars too, which is funny, because I didn't really expect a crossbar that serves memory to have much to do with distributing geometry at the front-end. Right? It made him even speculate that GK104 would only have 2x geometry, because that's what sensible people do.DarthShader said:So when are we going see the fixed bigK, aka GK100? Or at least GK110? GK112??
It all makes me wonder: what if Nvidia had used a 2x configuration instead of 4x, how much lower do you think the power of GK104 could have been?
What a missed opportunity...
Edit: I forgot the best one: the distributed geometry architecture is responsible for increase power consumption during compute. You can't make this up...
Last edited by a moderator:
After reading the release notes it was released not long ago. I think the previous version of that benchmark program is 2.5. I would like to see 3.0 vs 2.5 just to see if performance has changed or not using both cards.
Or you might simply ask Techpowerup's w1zzard if GPU-z 0.5.9 fully supports Kepler yet. [Hint: It doesn't, as you have noted above wrt to shaderclocks].
Yes, they allow AMDs higher pixel and texture fill as well as compute play to it's strength (and the front end and load distribution are not as much a stumbling block). In this case though, being already limited by tessellation performance, the normal advantage could very well be amplified quite a bit.
Last edited by a moderator:
This is curious btw
Notice the default clocks?
edit:
GPU-Z obviously detects "shaderclocks" wrong, likely just "new enough nvidia, let's use gpu clock * 2"
What I did notice, is that the gpu score shown here is 66% higher than my stock GTX 570! Wow!
If the performance difference remains the same in real games as well, I may have mislabeled the asking price as ridiculous.
Btw a Zotac 680 has been listed for 507 euros.
According to a fellow GTX 570 user, these are his results of 2.5 vs 3.0 with the 570 at stock.
Latency? Cross-bar interconnects are the first choice to be used in cases of low-latency communication between moderately large number of clients. The problem with complex cross-bar interconnects is the accumulation of hotspots due to signal crossings. In GF100 the distributed nature of both geometry processing (16x) and primitive setup (4x) asked for very dense wiring mesh. JHH said that this aspect of the architecture was the main reason the product delays and metal re-spins. The other obstacle was the large transistor leakage variance.
UniversalTruth
Veteran
There is already version 0.6.0 but I don't know how close you should be to TPU to actually get it.
