NVIDIA G-Sync

OpenGL OpenGL: Advanced Coding
1

System Config:-

1 HPxw9400 workstation.
2 GPUs (QuadroFX 4400G + QuadroFX 4500G)
1 G-Sync card (Primary -> 4500G & Secondary -> 4400G)
4 Monitors (2 in HorizontalSpan on 4500G & 2 in HS on 4400G)
WindowsXP + DualCore Intel CPU + 3GBRAM

From the NVIDIA Control Panel:-

The 4500G is configured as Master with internal framelock signal source. The 4400G is configured as Slave and synchronizes with the Master’s framelock signal.

Application code:-

2 OpenGL windows each one covering 1 horizontal span. So 2 (DC, RC)'s.
Enabled the 2 devices for G-Sync and checked its configuration.

Win1 -> HS1 -> 4500G -> Master -> Internal source -> Enabled
Win2 -> HS2 -> 4400G -> Slave -> External source -> Enabled

I reset the counter on the Master device. So far the setup looks fine.

When I render to the win1 & win2 and call swapbuffers on both sequentially it takes a while (about 15 sec) to display the frame and gets worse for subsequent frames. Application runs fine without G-Sync, no stall everything is smooth and real-time.

Are 4400G & 4500G compatible ? If so, is it possible to use both GPUs on the same machine as I’ve done ? Is there anything else that I’m missing in using the G-Sync ?

2

I don’t know anything about this.
However, I would seem more natural that the most powerful card should be the slave, as when the (slow) master finishes a frame, you can be sure that the slave have already finished its own.

You should use NV_swap_group apparently.

3

I’ve already tried switching the master and the slave but din’t make a difference. I’m using the swap group and swap barrier externsions. already.

4

if I remember correctly, there’s a 15 second time out with the nv swap group extension. So it sounds like one of your two window is waiting for the other to swap, or something.

5

Every other frame it takes approx. 14 sec to execute SwapBuffers of horizontal span 2. Have any idea why that is. Without G-Sync there is no such delay.

6

to the win1 & win2 and call swapbuffers on both sequentially

And what about calling swapbuffers in parallel ? knackered suggests that a 15 sec timeout means that one swap waits for the other to complete.