TDM and Optimus-like technology on desktop systems

[lowenz]

Maybe you don't know but the last iterations of Windows 10 on multi-GPU systems automatically choose the "best" GPU for the rendering (in D3D11 and lastest OpenGL applications) enabling the possibility of an Optimus-like configuration on desktop systems too: the active AND single-showed desktop on the monitor connected to the iGPU (the "integrated" GPU on the processor chip) where the output of the "real working" (typically dedicated, on the classic PCI-Express card) GPU is eventually routed thanks to the framebuffer being copied through the main System Memory.

The results (so you can see the actual overhead of the framebuffer copying/swapping process )

Pure Dedicated GPU (GeForce 1050 Ti):

"Optimus-Like" configuration (Intel HD 630 + GeForce 1050 Ti)

 

Not bad! Only a 9.5% of performance hit on a system with DDR4 2400 (so with better RAM modules and IMC you can reduce it to 5% I think)

Of course you can always force the rendering GPU through the dedicated interface in  "Display Settings -> Graphic Settings"

Edited May 17, 2020 by lowenz

[duzenko]

19 hours ago, lowenz said:

Maybe you don't know but the last iterations of Windows 10 on multi-GPU systems automatically choose the "best" GPU for the rendering (in D3D11 and lastest OpenGL applications) enabling the possibility of an Optimus-like configuration on desktop systems too: the active AND single-showed desktop on the monitor connected to the iGPU (the "integrated" GPU on the processor chip) where the output of the "real working" (typically dedicated, on the classic PCI-Express card) GPU is eventually routed thanks to the framebuffer being copied through the main System Memory.

The results (so you can see the actual overhead of the framebuffer copying/swapping process )

Pure Dedicated GPU (GeForce 1050 Ti):

"Optimus-Like" configuration (Intel HD 630 + GeForce 1050 Ti)

 

Not bad! Only a 9.5% of performance hit on a system with DDR4 2400 (so with better RAM modules and IMC you can reduce it to 5% I think)

Of course you can always force the rendering GPU through the dedicated interface in  "Display Settings -> Graphic Settings"

Good to hear things are moving in that direction but 10% (1.5msec) is un-explainable slow

1080p framebuffer is just 8MB of data. That means the effective transfer speed is ~5GB/sec. Meh!

Compare to promised ~15GB/sec for PCIEx16 3.0. Unless, of course, you're running in x8 or 2.0 mode.

[lowenz]

3.0 x16 here!

But DDR4 2400 only.

[duzenko]

2 hours ago, lowenz said:

3.0 x16 here!

But DDR4 2400 only.

Dual channel? That's what 40GB/sec?

[lowenz]

Dual channel yes, but the limitation resides in the IMC and PCI-E management as you say. Still more speed (latencies are already low in my system) in the memory section could gives more clear results.

[lowenz]

I can confirm all is working as intended !

I've literally unplugged the GeForce from the monitor input but I'm getting some solid 55-56 FPS in TDM  (in the scene above)

With NO wizardry in Windows at all, it's just there.

Edited May 19, 2020 by lowenz

[duzenko]

Does not work the other way around

When monitor is plugged to the nVidia, it won't use the IGP even though configured in Settings

Job half-done, as always with M$

[lowenz]

Good old Source (SDK 2013 Single Player Upcoming branch), all maxed out (MSAA 4x too), with 999 FPS cap

Pure Dedicated GPU (Radeon RX 570 / 2050 MHz VRAM) -> 548 FPS

Optimus-Like configuration (Intel HD 630 + Radeon RX 570/ 2050 MHz VRAM ) -> 379 FPS

 

So the overhead is (548-379)/548= 31%

Edited May 26, 2020 by lowenz

[lowenz]

With modern engine like the Unreal Engine 4 in Borderlands 3 the overhead is 0 (yes, ZERO)

So maybe the main issue is the texture streaming optimization missing in old engines?

 

Borderlands 3 has the same average framerate in the integrated benchmark!

Edited May 26, 2020 by lowenz