I do think it's really difficult to hold back from setting up a multi-GPU rig. You see all the benchmark results and mods of the month and it just looks too cool not to do it.
However, unless you are running max settings on one very large screen or a mutli screen then 2,3 or more GPU's is kind of pointless in today's gaming world. Folding is another matter all together.
I'd really like to but I can't really see the immediate benefit until developers fully utilise the potential but they aren't making games assuming every has multi GPU's.
Until the 'leading' console has some sort of multi-processor (or maybe all the consoles?) SLI will continue to be the farce it has been since voodoo2 actually pulled it off... can't get my head around why it is so hard for companies/devs to get behind the drivers and make SLI something other than Energy companies income and proof that you have more money than sense...
Ive always been from the school of buy a faster single card over a couple of slower 1s, and this hasnt changed my mind.
The problem with current multi GPU setups is that performance is so hit and miss, Yes in some games you may get a massive boost in performance but as the charts show in some it can cripple your system meaning you have to disable a GPU, which imo defeats the point in having 2 in the first place.
Im no coder, nor an expert on the technology used but surely it cant be that hard to produce drivers that allow 100% scaling across the board? Or is it down to how individual games are coded and no amount of driver updates will ever make a difference?
Originally Posted by lysol i still have no idea why we don't have multicore gpus by now. if we had multicore gpus we wouldn't have to muck about with multiple cards.
Its the size of the Silicon. Graphics chips are huge beasts. Every time you add more transistors, you increase the surface area for defects.
These driver updates seems to be rather workaround than solution for real multi GPU support.
I wonder whether it's possible to design graphic API(D3D or OGL) with multi-GPU support so that it's no longer issue of drivers to get decent fps.
Just specced a 5770 for a friend - very pleasantly surprised with the performance for a single fairly midrange device. Loaded up Far Cry 2, wound all the sliders up as far as they'll go at 1920x1200, and it's very nearly perfect.
He could in theory stick a second 5770 in there - the board supports it - but even before reading this article I advised not to bother. It's really pretty depressing that they can't do any better than this. I mean, what's the fricken hold-up? Render one vertical half of the screen on each board and have done with it, because it's abundantly clear that whatever cleverer technique they're trying for doesn't work very reliably.
I've owned dual GPU's since the days of 3DFX (God I miss those days, Carmageddon TDM was such fun), such cards as ATI X1900XT's and Nvidia 8800GTX's and in every case except 3DFX I've found dual card performance to be very disappointing.
I'm sticking to single cards for the time being... I have lost too much money chasing those frame rates that while theoretical, in practice are often impossible to achieve.
I could understand the need fro dual card setups when it first started, as the PC was the lead platform for cutting edge development of new and exciting graphics engines and therefore every extra frame counted.
But now?.....when all the Pc gets (with the odd exception) are console engine ports, based on 2005 level specs......
Even my budget Pc costing £400, a Phenom II 550 and a 4850, can run any console port at twice the resolution and at least a 50% better frame rate...as a 360 or PS3, including the latest ones like Resident Evil 5 and Batman Arkham Asylum.
Originally Posted by Phil Rhodes I mean, what's the fricken hold-up? Render one vertical half of the screen on each board and have done with it, because it's abundantly clear that whatever cleverer technique they're trying for doesn't work very reliably.
Doesn't help much though as you could find that all the complex stuff is on one half of the screen, leaving one card struggling. Sure, you'll still see a performance increase across the board in all games, but it'd be far from the holy grail of 100% scaling. You'd be better of splitting the frame into small tiles and having cards working on them alternately. Or maybe just have the card that is available first take the next tile in line (no need for them to do an even number after all, otherwise run into the same problem above). Of course there will be a trade off between tile size and overheads, but I'm sure if you get enough top engineers together they could work out the optimal number!
I quite like the look of the xfire 5770s. I assume you can disable one for games where there is a performance hit?
It'd have been good to see some overclocking on the cards, just to see how the extra heat impacts it and what extra performance could be squeezed out!
Originally Posted by eek Doesn't help much though as you could find that all the complex stuff is on one half of the screen, leaving one card struggling. Sure, you'll still see a performance increase across the board in all games, but it'd be far from the holy grail of 100% scaling. You'd be better of splitting the frame into small tiles and having cards working on them alternately. Or maybe just have the card that is available first take the next tile in line (no need for them to do an even number after all, otherwise run into the same problem above). Of course there will be a trade off between tile size and overheads, but I'm sure if you get enough top engineers together they could work out the optimal number!
So something like mosaic SLI (which nVidia offers for Quadro cards, but not consumer cards) or the original approach to multi-card display - Scan Line Interleaving... one card takes the odd lines, the other card takes the even ones. Would get a little more complicated with tri- or quad- setups, mind.
I'd like to see Tri-Fire EyeFinity where one card handles each screen, but ATi said it's not on their list of things to do. :(
do you not think you're being a little greedy? I say that because even theoretically 100% scaling is impossible is it not? Even going from 1 cpu to 2 doesn't have 100%scaling more like 80-90%. While i agree about dawn of war and call of duty every other test ran gave imo perfectly acceptable results. Why is dawn of was like that? Was it something to do with a recent patch?
Maybe should compare the items and their costs too. I mean you can cheaply make tri-sli with 5770!!! and still worth as much as a 5870 or GTX 295! What's the speeds then? Maybe on an overclocked Q6600 at 3.2?
Originally Posted by Toka Just out of interest Bit-Tech...
Why do you bother with 12x10 and 16x10 resolutions in tests like this. Or have you brainwashed Harry into enjoying the benchmarking process?
I bet thats it isnt it - you snuck one of those hypnosis tapes under his pillow..
Because people do game at these resolutions. And unfortunately not everyone realises that to make the most out of an amazing card or pair of cards you need to pump up the resolution. So really they are doing the less educated among us a service by letting them know if you do game at 600x480 then sli or crossfire should be the lowest down on your upgrade list.
Also it serves as a good comparison against the higher resolutions.
Hoary old gamers will remember that SLI originally stood for scanline interleave, which meant that each board was rendering alternate lines. Now, this does mean that you have to maintain the entire geometry for the scene on both boards and you can't do any sort of fustrum-based optimisation, but then I don't see how you could do that on the sort of variable-split approaches that are currently being used. Two adjacent rows of pixels are clearly likely to have very similar complexity.
I presume there is some implementation problem with doing this of which I'm not aware, because otherwise it makes perfect sense.
The new 5870's shine at higher res, It seems abit dull that you don't review this area. I know most people won't use this type of res but it does still show how well it will perform.
Do benchmarks @ 5040x1050 and 5760x1200 and then you will see how well these cards perform.
+1 for 5040x1050 and 5760x1200. ATI are pushing insane resolutions, it would nice to see if a single card can play these and their CF settup will provide any benifit. I know that the only reason ill upgrade from HD4870 is to play at these resolutions. already building towards it by buying a 2nd monitor.
Originally Posted by storm20200 do you not think you're being a little greedy? I say that because even theoretically 100% scaling is impossible is it not? Even going from 1 cpu to 2 doesn't have 100%scaling more like 80-90%. While i agree about dawn of war and call of duty every other test ran gave imo perfectly acceptable results. Why is dawn of was like that? Was it something to do with a recent patch?
Anything above 50% is pretty impresive improvement(even though it doesn't sound like much). Reflections, all sorts of shaders, dynamic lights(shadows) etc... - there'll always be certain amount of processed data required by both GPUs to render "their" part of screen. It's definitely not as easy as splitting screen in halves or lines. That's why I think graphic API with multi GPU rendering in mind might reduce unnecessary re-calculations or even dynamicaly divide screen regions between GPUs.
I doubt we will see 100% improvement like in case of old 3Dfx cards. Quake-age is the past, just like static lights and flat surfaces...
Comments 1 to 25 of 58
ReplyHowever, unless you are running max settings on one very large screen or a mutli screen then 2,3 or more GPU's is kind of pointless in today's gaming world. Folding is another matter all together.
I'd really like to but I can't really see the immediate benefit until developers fully utilise the potential but they aren't making games assuming every has multi GPU's.
(Sorry, I'm bias towards gaming rigs)
The problem with current multi GPU setups is that performance is so hit and miss, Yes in some games you may get a massive boost in performance but as the charts show in some it can cripple your system meaning you have to disable a GPU, which imo defeats the point in having 2 in the first place.
Im no coder, nor an expert on the technology used but surely it cant be that hard to produce drivers that allow 100% scaling across the board? Or is it down to how individual games are coded and no amount of driver updates will ever make a difference?
EDIT: Also please for the love of god fix the phenom 2 550 review that says it has 8MB of L3 cache on the first page T_T
Its the size of the Silicon. Graphics chips are huge beasts. Every time you add more transistors, you increase the surface area for defects.
5770 is impressive in some games
I wonder whether it's possible to design graphic API(D3D or OGL) with multi-GPU support so that it's no longer issue of drivers to get decent fps.
Why do you bother with 12x10 and 16x10 resolutions in tests like this. Or have you brainwashed Harry into enjoying the benchmarking process?
I bet thats it isnt it - you snuck one of those hypnosis tapes under his pillow..
He could in theory stick a second 5770 in there - the board supports it - but even before reading this article I advised not to bother. It's really pretty depressing that they can't do any better than this. I mean, what's the fricken hold-up? Render one vertical half of the screen on each board and have done with it, because it's abundantly clear that whatever cleverer technique they're trying for doesn't work very reliably.
Still none I'm afraid. I believe ATI is waiting for Stanford to release a new version of the client that'll run on OpenCL:
http://folding.typepad.com/news/2009/09/update-on-new-fah-cores-and-clients.html
I'm sticking to single cards for the time being... I have lost too much money chasing those frame rates that while theoretical, in practice are often impossible to achieve.
But now?.....when all the Pc gets (with the odd exception) are console engine ports, based on 2005 level specs......
Even my budget Pc costing £400, a Phenom II 550 and a 4850, can run any console port at twice the resolution and at least a 50% better frame rate...as a 360 or PS3, including the latest ones like Resident Evil 5 and Batman Arkham Asylum.
Dual cards are just so utterly pointless.
I quite like the look of the xfire 5770s. I assume you can disable one for games where there is a performance hit?
It'd have been good to see some overclocking on the cards, just to see how the extra heat impacts it and what extra performance could be squeezed out!
So something like mosaic SLI (which nVidia offers for Quadro cards, but not consumer cards) or the original approach to multi-card display - Scan Line Interleaving... one card takes the odd lines, the other card takes the even ones. Would get a little more complicated with tri- or quad- setups, mind.
I'd like to see Tri-Fire EyeFinity where one card handles each screen, but ATi said it's not on their list of things to do. :(
Because people do game at these resolutions. And unfortunately not everyone realises that to make the most out of an amazing card or pair of cards you need to pump up the resolution. So really they are doing the less educated among us a service by letting them know if you do game at 600x480 then sli or crossfire should be the lowest down on your upgrade list.
Also it serves as a good comparison against the higher resolutions.
Hoary old gamers will remember that SLI originally stood for scanline interleave, which meant that each board was rendering alternate lines. Now, this does mean that you have to maintain the entire geometry for the scene on both boards and you can't do any sort of fustrum-based optimisation, but then I don't see how you could do that on the sort of variable-split approaches that are currently being used. Two adjacent rows of pixels are clearly likely to have very similar complexity.
I presume there is some implementation problem with doing this of which I'm not aware, because otherwise it makes perfect sense.
Do benchmarks @ 5040x1050 and 5760x1200 and then you will see how well these cards perform.
Anything above 50% is pretty impresive improvement(even though it doesn't sound like much). Reflections, all sorts of shaders, dynamic lights(shadows) etc... - there'll always be certain amount of processed data required by both GPUs to render "their" part of screen. It's definitely not as easy as splitting screen in halves or lines. That's why I think graphic API with multi GPU rendering in mind might reduce unnecessary re-calculations or even dynamicaly divide screen regions between GPUs.
I doubt we will see 100% improvement like in case of old 3Dfx cards. Quake-age is the past, just like static lights and flat surfaces...
-
« Previous
-
1
-
2
-
3
-
Next »
Discuss in the forums