harmattan
Supreme [H]ardness
- Joined
- Feb 11, 2008
- Messages
- 5,129
NoQuarter over at Anand has a good hypothosis on what the delay/problem is. If this is the case, we may be looking at a hardware/bandwidth limitation that will prevent 3D Surround from ever performing at higher resolutions...
My guess was something to do with SLI bandwidth and sending the frame buffer alternating directions each frame. SLI normally functions in AFR by sending the frame from the slave card to the master card for the frame the slave card is responsible for.
Supposedly the SLI link is capable of 1GB/s. Nvidia already stated (according to EVGA forums) that its Surround implementation would be limited to 5760x1080 max (ie can't use 1920x1200 monitors). The reason this would be is because the bandwidth required to send that frame buffer is too much.
ATI already said the Crossfire link is barely enough to handle Eyefinity resolutions and it's likely some frames get sent via PCI-e bus when the Crossfire link is saturated causing the stuttering a few users have reported when running Crossfire Eyefinity. The Crossfire link is capable of 0.9GB/s but you can double them up between your 2 cards, providing 1.8GB/s.
If the SLI link is capable of 1GB/s and a 5760x1080 resolution is 6220800 pixels, and each pixel is 4 bytes of data, the frame buffer would be 24883200 bytes or 0.023174286 Gigabytes. If this is all true, the max frame rate transfer between cards would be 43 frames per second not including necessary overhead but assuming it has to send the full frame.
In Eyefinity's implementation the frame buffer only needs to be sent across the Crossfire link every other frame (when the slave card is on duty). In Surround's implementation the frame buffer or at least a portion of it needs to be sent across the SLI link every frame (from master to slave, then slave to master).
If the drivers are capable of it it could be possible to send 2/3 of the frame buffer from slave to master for one frame, then 1/3 of the frame buffer from master to slave the next frame, cutting the bandwidth required for frame copies in half (sort of).
The SLI bridge apparently was designed to be uni-directional also, which may cause some latency issues waiting for the bridge to clear up to send data the other way.
Also using SFR rendering method gets terribly complicated as both cards would be rendering parts of the frame the other card needs.
My guess was something to do with SLI bandwidth and sending the frame buffer alternating directions each frame. SLI normally functions in AFR by sending the frame from the slave card to the master card for the frame the slave card is responsible for.
Supposedly the SLI link is capable of 1GB/s. Nvidia already stated (according to EVGA forums) that its Surround implementation would be limited to 5760x1080 max (ie can't use 1920x1200 monitors). The reason this would be is because the bandwidth required to send that frame buffer is too much.
ATI already said the Crossfire link is barely enough to handle Eyefinity resolutions and it's likely some frames get sent via PCI-e bus when the Crossfire link is saturated causing the stuttering a few users have reported when running Crossfire Eyefinity. The Crossfire link is capable of 0.9GB/s but you can double them up between your 2 cards, providing 1.8GB/s.
If the SLI link is capable of 1GB/s and a 5760x1080 resolution is 6220800 pixels, and each pixel is 4 bytes of data, the frame buffer would be 24883200 bytes or 0.023174286 Gigabytes. If this is all true, the max frame rate transfer between cards would be 43 frames per second not including necessary overhead but assuming it has to send the full frame.
In Eyefinity's implementation the frame buffer only needs to be sent across the Crossfire link every other frame (when the slave card is on duty). In Surround's implementation the frame buffer or at least a portion of it needs to be sent across the SLI link every frame (from master to slave, then slave to master).
If the drivers are capable of it it could be possible to send 2/3 of the frame buffer from slave to master for one frame, then 1/3 of the frame buffer from master to slave the next frame, cutting the bandwidth required for frame copies in half (sort of).
The SLI bridge apparently was designed to be uni-directional also, which may cause some latency issues waiting for the bridge to clear up to send data the other way.
Also using SFR rendering method gets terribly complicated as both cards would be rendering parts of the frame the other card needs.