Overclock.net › Forums › Industry News › Hardware News › [CNET] AMD Bests Nvidia With Graphics Strategy
New Posts  All Forums:Forum Nav:

[CNET] AMD Bests Nvidia With Graphics Strategy - Page 3

post #21 of 43
ATi.........................I love you.


so.......beautiful........


But finally I picked the right motherboard to stick with for a while! hurray for CF chipsets. Go x48 and ATi
post #22 of 43
Thread Starter 
Quote:
Originally Posted by Nubcake View Post
The interconnect thingy, is it hardware based or driver based?
The proprietary interconnect for the 4870X2 is a hardware thing since it will be put on the board. However, it also seems as if it's fairly reliant on software (ie drivers) to function correctly, at least according to Fudzilla:

Quote:
The dual RV770-based R700 card is supposed to launch in August time, rather late than early, and the hardware is ready; it is now all up to the drivers.
post #23 of 43
but if the scaling is done internally, why would the drivers be needed to see it? It would just happen without the driver telling it to in theory. That also makes me wonder if it would happen right on an nvidia chipset. :?
post #24 of 43
Last i checked Nvidia was too busy Re-labeling the 8800GT(s) cards as 9800GTX+.
post #25 of 43
Thread Starter 
Quote:
Originally Posted by slngsht View Post
but if the scaling is done internally, why would the drivers be needed to see it? It would just happen without the driver telling it to in theory.
It's pure speculation at this point, but here's my theory:

In its current state, CrossFire uses Alternate Frame Rendering to scale performance accross multiple GPUs. Essentially what it does is renders Frame A on one GPU, but while Frame A is being rendered the driver is also sending Frame B to the second GPU. This way by the time Frame A is completed, Frame B will already be partially done (ideally it will be about halfway or better). Obviously this happens very quickly, and is repeated over and over again alternating which GPU it's using. This is why adding more than a second GPU has diminishing returns since it can only render so far ahead.

In order to make the 2 GPUs completely transparent to the driver/OS, the hardware would have to handle the AFR. However, this probably wouldn't be practical since AFR is a software solution. The driver plays a very key role because it will essentially "lie" to a game about when Frame A is done rendering so that the game will let the driver start rendering Frame B ahead of time.

For a true, transparent multiGPU solution, they would have to use a different method that would make sense to do in hardware. Two possible choices would be Split-Frame Rendering or SuperTiling. SFR is exactly what it sounds like, one GPU would do the top half, and the other would do the bottom. The problem with this is in situations where one half of the screen is a lot more complex than the other half. When this happens, it won't be rendered very much faster than a single GPU, because it would still have a single one doing all the "hard" stuff. SuperTiling is a similar concept but involves breaking the frame into tiles rather than just splitting it in half.

However, in order for either SFR or SuperTiling to achieve the kind of scaling that ATI is talking about here, they would have had to develop some kind of load balancing algorithm so that each GPU is performing the same amount of work. This is difficult enough to even pull off in software, and I would think that a hardware solution would be even harder. I don't think that ATI is quite at the point where they can have this happen independently of the drivers. If the 4870X2 is being delayed due to drivers, that leads me to believe that they are doing something different from AFR but it's not done in hardware at the moment.

Quote:
That also makes me wonder if it would happen right on an nvidia chipset. :?
It should still work fine, since it's not even a CrossFire connection anymore.
post #26 of 43
So technically this chip will be faster than two 4870's in CrossFire? If this card is cheaper than the 280GTX, nVidia is ruined this time around.
00-Raiser
(13 items)
 
  
CPUMotherboardGraphicsRAM
Core i7 920@3.4GHz ASUS P6T X58 Sapphire ATI 4870 512MB Gskill 6GB 1333MHz 
Hard DriveOptical DriveOSMonitor
2x640GB WD Black (RAID 0) 22x Samsung DVD Burner Windows Server 2008: Enterprise Edition Dell S2309W 
KeyboardPowerCaseMouse
Logitech G15 Corsair 850TX Cooler Master Cosmos 1100 Logitech G9 
  hide details  
Reply
00-Raiser
(13 items)
 
  
CPUMotherboardGraphicsRAM
Core i7 920@3.4GHz ASUS P6T X58 Sapphire ATI 4870 512MB Gskill 6GB 1333MHz 
Hard DriveOptical DriveOSMonitor
2x640GB WD Black (RAID 0) 22x Samsung DVD Burner Windows Server 2008: Enterprise Edition Dell S2309W 
KeyboardPowerCaseMouse
Logitech G15 Corsair 850TX Cooler Master Cosmos 1100 Logitech G9 
  hide details  
Reply
post #27 of 43
Quote:
Originally Posted by onexero View Post
So technically this chip will be faster than two 4870's in CrossFire? If this card is cheaper than the 280GTX, nVidia is ruined this time around.
Much.

Edit: And darkcloud man... If they can genuinely trick the computer into thinking it is 1GPU, they wont even have to work on an advanced form of super4tiling or SFR, both gpus can just work on different parts of the same frame, because if it works the way I think its going to, its basically going to be like 1 enormous die, with some of the shaders doing water, others doing spec mapping, etc.

I think the 20% drop in performance, as opposed to a full 100% increase from adding a second die will come from the latency between the two cores, which if they are talking through the 3600mhz memory like some of the rumors say, really shouldn't even be that high
Edited by HugeDink - 6/29/08 at 1:01pm
Von Strauss
(13 items)
 
  
CPUMotherboardGraphicsRAM
Athlon 5000x2 BE @ 3.2 ASUS M2R-32 MVP Sapphire HD4850 700/1100 2x1gb G.Skill HZs 
Hard DriveOptical DriveOSMonitor
250gb WD IDE lol LG DVD-RAM w/ Lightscribe Vista Ultimate 32-bit Sceptre, 1280x1024, 5ms 
PowerCase
Corsair TX 650W Enermax Chakra w/ 250mm Side Panel Fan 
  hide details  
Reply
Von Strauss
(13 items)
 
  
CPUMotherboardGraphicsRAM
Athlon 5000x2 BE @ 3.2 ASUS M2R-32 MVP Sapphire HD4850 700/1100 2x1gb G.Skill HZs 
Hard DriveOptical DriveOSMonitor
250gb WD IDE lol LG DVD-RAM w/ Lightscribe Vista Ultimate 32-bit Sceptre, 1280x1024, 5ms 
PowerCase
Corsair TX 650W Enermax Chakra w/ 250mm Side Panel Fan 
  hide details  
Reply
post #28 of 43
The basic math is wrong. 1.8+1.8 is = 3.6 not 2.5.
post #29 of 43
Thread Starter 
Quote:
Originally Posted by 45nm View Post
The basic math is wrong. 1.8+1.8 is = 3.6 not 2.5.
It doesn't keep the same scaling from 1-2 GPUs as it does from 2->4....
post #30 of 43
Quote:
Originally Posted by 45nm View Post
The basic math is wrong. 1.8+1.8 is = 3.6 not 2.5.
no, your equation is wrong.

I'm really not even sure what your equation is even trying to say.

like dark already said, since a 4870 is about 10% slower than a GTX280 on average, so, if we say that the GTX280 is our constant, at 100, then a 4870 is an 90, so two of these in the proposed 4870X2 setup would be 90x1.8, which equals 162, meaning that an HD4870X2 should be about 62% faster than a GTX280.
Von Strauss
(13 items)
 
  
CPUMotherboardGraphicsRAM
Athlon 5000x2 BE @ 3.2 ASUS M2R-32 MVP Sapphire HD4850 700/1100 2x1gb G.Skill HZs 
Hard DriveOptical DriveOSMonitor
250gb WD IDE lol LG DVD-RAM w/ Lightscribe Vista Ultimate 32-bit Sceptre, 1280x1024, 5ms 
PowerCase
Corsair TX 650W Enermax Chakra w/ 250mm Side Panel Fan 
  hide details  
Reply
Von Strauss
(13 items)
 
  
CPUMotherboardGraphicsRAM
Athlon 5000x2 BE @ 3.2 ASUS M2R-32 MVP Sapphire HD4850 700/1100 2x1gb G.Skill HZs 
Hard DriveOptical DriveOSMonitor
250gb WD IDE lol LG DVD-RAM w/ Lightscribe Vista Ultimate 32-bit Sceptre, 1280x1024, 5ms 
PowerCase
Corsair TX 650W Enermax Chakra w/ 250mm Side Panel Fan 
  hide details  
Reply
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Hardware News
Overclock.net › Forums › Industry News › Hardware News › [CNET] AMD Bests Nvidia With Graphics Strategy