I game at 2560 so for me its more important than other res, but its true 1920x are far more popular, and 1440/1680x even more so. Everyone has to choose whats best for them. In any case, considering GTX580 is 10-15% faster on 1680 and 1920 with 8x AA, while costs 140$ more and thats with 6970 ~beta drivers for a new arch. I'm guessing 6970 will catch up Nvidias top dog even more, while NV will have to drop prices (once they can supply). Win-win for customers, regardless which brand they choose.
AMD: R9xx Speculation
- Thread starter Lukfi
- Start date
TechPowerUp for example doesn't use 8xAA, yet at 2560x1600 GTX580 averages 10% faster than 6970, while at 1920x1200 it's 13% faster.
http://techpowerup.com/reviews/HIS/Radeon_HD_6970/29.html
At Computerbase it looks even better for 6970:
http://www.computerbase.de/artikel/...eon-hd-6970-und-hd-6950/26/#abschnitt_ratings
1920 No AA GTX580 wins by nearly 12%
1920 4xAA GTX580 wins by a bit over 15%
1920 8xAA GTX580 wins by nearly 15%
2560x166 however is another story
2560 No AA GTX580 wins by just a bit over 5%
2560 4xAA GTX580 wins by nearly 12%
2560 8xAA GTX580 wins by just over 2%
What's curious is how much stronger the 4xAA is on nV compared to AMD/ATI
I've been thinking. The reviews say the reason for VLIW4 is that they were seeing 3.4 unit utilization. So, the fundamental design goal of Cayman was to optimize the GPU. This makes me think that drivers are probably about as efficient as they are going to get. It's not so much a new architecture as it is a refined one. Opinions?
Also, considering the lag time in design, the 3.4 unit utilization determination would have been in 2007-8? Could that no longer be optimal?
Also, considering the lag time in design, the 3.4 unit utilization determination would have been in 2007-8? Could that no longer be optimal?
While it is possible that the VLIW4 shaders are scheduled well, that is only one part of the whole design. A driver could perfectly schedule slots within the ALU clauses and then fail to yield an improvement because this reduced the occupancy of the SIMDs below the point that they could hide latency, or did a poorer job with shared memory or geometry processing.
VLIW4 does not help with being "front-end limited", whatever components that actually covers.
It would make the front end part of the process look worse, since an improved back-end will mean that a larger share of time is taken up by the unimproved parts of the graphics pipeline.
The rest of the chip did make some changes, but they were also modest refinements.
There is the more flexible concurrent scheduling of kernels, but that is something that may require software that targets it.
The LDS no longer consumes ALU instructions to load in data, but it may not have a full effect until code is written to more heavily rely on it, since up till now it was much less of a win because the register file is so large and the FLOPs hit could be significant.
The memory pipeline is still very specialized, though it had some efficiency changes. I would expect at some point AMD is going to need to close the loop so that there is a more generalized read-write path.
I'm curious about the components of the 40 cycle clause change latency. Couldn't this be improved?
Is a significant portion of the latency is due to building the clause context?
Could this be improved in the case of successive clauses for the same shader, which would in theory be continuations of the same instruction stream?
VLIW4 does not help with being "front-end limited", whatever components that actually covers.
It would make the front end part of the process look worse, since an improved back-end will mean that a larger share of time is taken up by the unimproved parts of the graphics pipeline.
The rest of the chip did make some changes, but they were also modest refinements.
There is the more flexible concurrent scheduling of kernels, but that is something that may require software that targets it.
The LDS no longer consumes ALU instructions to load in data, but it may not have a full effect until code is written to more heavily rely on it, since up till now it was much less of a win because the register file is so large and the FLOPs hit could be significant.
The memory pipeline is still very specialized, though it had some efficiency changes. I would expect at some point AMD is going to need to close the loop so that there is a more generalized read-write path.
I'm curious about the components of the 40 cycle clause change latency. Couldn't this be improved?
Is a significant portion of the latency is due to building the clause context?
Could this be improved in the case of successive clauses for the same shader, which would in theory be continuations of the same instruction stream?
3.4 unit utilization would be barely better what they got with VLIW5, so I doubt it's as good as it's going to get.
Also, utilization of VLIW-cores isn't the only factor related to how fast the drivers can get the cards to run
I averaged Anandtech's games at 1920X1200 (figured it was closest to 1080P, and that 1080P is probably the most common resolution) and 6970 was 6% slower...than GTX570, forget 580.
I think it would have been at least 25% than 580. There were a couple games Nvidia excelled in, those were only ~20 faster on 570, but ~40+% faster on 580, so the averages would have really gotten ugly.
Anandtech used older driver, and got far worse results than those who used few days newer driver, I guess it answers the question if drivers can help Caymans to look better
Because for whatever reason, some people still insist drivers are as good as it gonna get And even that "newer driver" is still in beta stage, as usually AMD focuses on stability first, speed comes later. If 6970 is already close to GTX580, its not hard to imagine what happens later when drivers mature. And 6970 isnt even meant to face 580, its 6990 job. Tough year for Nvidia, again 
This resolution is common for average user, but is it also common for high-end user (~5%)? I'm not convinced, that high-end user, who spends $400-$1000 per year for graphics card(s) will use it with $150-$250 1080p LCD…
Exactly my point, and while its true some may be using high-end cards with 1920 res or lower, but ~5% are enthusiasts for a reason. LCDs serve longer than GPUs, so spending ~600$ per year on high-end cards and not spending ~1-1.5k $ on LCDs every 3 years (?), doesnt make much sense. Regardless if its 30", or three 24", you dont really want to play with GTX580/6990, etc. on a single 1920 or lower. Or you could just save the money and get cheaper GPU and spend money elsewhere
Blazkowicz
Legend
well I'm sick of seeing $700 PCs with $5 keyboard and $5 speakers, but people do it all the time.
you might have nonetheless "reasonable" high end buyers who don't junk their expensive as fuck (i.e. 400€ or more) 1200p or 1080p IPS panel only because it's one or a few years old. ditto for GPUs you might as well skip each other generation or half-generation.
on another note, 120Hz refresh will make the framerates more useful again.
if I do some 100Hz gaming on not too heavy games that run very fast, I call 60fps a framerate drop sometimes.
Last edited by a moderator:
DavidGraham
Veteran
I don't think the new driver added any kind of performance improvement for the HD 6900 just yet , otherwise it would have been written in the release notes .Anandtech used older driver, and got far worse results than those who used few days newer driver, I guess it answers the question if drivers can help Caymans to look better
I don't think NVIDIA is in any troubled position this year at all , at least relative to last year or Q2 2010 , they have solid offerings at the right price points , and they have their reputation , it will be a close battle .
Highly unlikely , AMD archeticture is just better at pixel processing due to their high shader count , and fill rate .. Increasing display resolution shifts the load to better utilize these advantages .I was thinking about it the other day, and my guess is, Nvidia has undiscriminating brute force approach (need it or not, here comes 200FPS on your 1024x.. LCD!"
IMO , increasing the resolution isn't the optimum solution for increasing realism and visual impact in games .. detailed textures , lighting and primitives count are just as important if not more .
4 of my friends are enthusiasts (HD 5970 , 5850/5870 CF , GTX 480 SLI) , and all of them use 1920x1080 displays , that's because 4 megapixel displays are rare commodities , expensive and heavy , hard to sell when upgrading and always need very good hardware to handle heavy games , and believe me , sometimes even crossfire and SLI fail to provide consistent performance at this resolution .
Last edited by a moderator:
I have both a 30" 2560x1600 and a 24" 1920x1200 screen and game on the latter since getting adequate performance on the 30" is hit or miss and the 24" is a better overall display. I would be surprised if 5% of the people playing games are doing so at 30" or multi-display resolutions. Having said that, AMD's cards do close the gap dramatically as resolution increases. It's a new phenomenon so it could be due to a deficiency in Fermi's data paths or some genius in the EG/NI archs.
DavidGraham
Veteran
I reckon it's the former , GTX 580/70 cards just can't utilize more than 32 ROPs during pixel fill rate operations , only during blending does the increased ROPs count pay off , and I think that is the reason NVIDIA has an advantage at 4X AA .
I'm not saying, that 5% of people use 30" displays / multi-display configuration, but enthusiasts buing the most expensive hardware use it more often, than common users, so hi-res performance shouldn't be ignored.
There are a few LCD models between 1920×1200 and 4MP displays, e.g. the 2560×1440 DELL, which is also less expensive than typical 4MP displays. I think that if a particular game doesn't run well at full 2560×1600, you can run it in lower-resolution (non-interpolated, of course)…
There are a few LCD models between 1920×1200 and 4MP displays, e.g. the 2560×1440 DELL, which is also less expensive than typical 4MP displays. I think that if a particular game doesn't run well at full 2560×1600, you can run it in lower-resolution (non-interpolated, of course)…
So you cant even imagine how drivers revisions add extra performance to the new arch?
Oh, and why AMD would mention in notes perf differences between one pre-release revision and another? There were floating several same number revisions with different performance :smile: Close battle where? In high-end NV already lost, its barely faster than much cheaper 6970, not really faster than a year old 5970, and is preparing to be spanked by 6990
NV is in slightly better position, but nothing changed much, NV still will be losing market share.DavidGraham
Veteran
I never said AMD can't extract more performance in future drivers , on the contrary I believe this will happen for sure .So you cant even imagine how drivers revisions add extra performance to the new arch?
As for release notes , they ALWAYS contained how much performance has improved for each title , this is a trend for both companies , if the driver contained general improvements across a wide range of titles , it will also be mentioned in the notes.
See the above posts , "barely faster" is a description for HD 6970 in comparison to GTX 570 , extreme enthusiasts with 4MP displays opt for CF or SLI solutions , not single cards , so GTX 580 is still a valid choice , especially for those who want to avoid the draw-backs of dual GPU solutions .Close battle where? In high-end NV already lost, its barely faster than much cheaper 6970, not really faster than a year old 5970, and is preparing to be spanked by 6990
GTX 570 is in even better position , and the incoming GTX 560 will stir things up for sure ..
BTW , @ 2500x1600 4XAA , a GTX 580 is on equal footing with an HD 5970 :
http://www.computerbase.de/artikel/...-570/20/#abschnitt_performancerating_mit_aaaf
Only future drivers? AMD cant add extra performance now? How many times we have to repeat obvious examples of review sites using diff revisions and getting diff results?
No they dont, not with pre-release drivers like its now.
Since 570/580 are close in performance, its obvious 6970 as in between card will be barely faster then one and barely slower than another
Also what do you mean "extreme enthusiasts with 4MP displays opt for CF or SLI solutions , not single cards , so GTX 580 is still a valid choice , especially for those who want to avoid the draw-backs of dual GPU solutions", so only dual GPU cards have draw-backs, but CF/SLi dont? BTW, sorry to disappoint you, but even lowly and cheap 6850 CF beats 580 SLi in ultra resolutions, what to speak of Caymans...2% difference doesnt register at all to you as close cards in one case, but in another its equal footing? You rejected 2560 in the first place, why you even quoting it?
Similar threads
