Shared GPU memory stopping RTX2070 super being used in VR

User Support Hub Virtual Reality (VR)
1

I have spent a few days now trying to get VR working reasonably on my RTX2070 super. My son also has a system with a RTX 3090 in it which is running very well in VR (just a few odd stutters here and there). I am using an HP Reverb G2

I read a few days ago this article

MS Flight Simulator 2020 Benchmarks: CPU, GPU, Memory Performance (premiumbuilds.com)

which went a long way to explaining how to get a balanced CPU/GPU setup. In particular it was very interesting to read about how counter intuitive it can be to sometimes increase the load on the GPU to get a more fluid experience. Based on this I tried to fine tune the 3090 system and eventually ended up with a good set of settings. While doing this I noticed in task manager that the memory available to graphics was up at 40GB! The card only has 24GB but the system was also making 16GB of main system ram available for use by the card if needed. Of course with 24GB the card never needs more than this and typically in non VR mode at Ultra settings uses less than 8GB which is the spec MS says is good for high end systems. However, when you go to VR mode this jumps to about 12GB depending on your settings. On the 3090 this is all allocated to the VRAM on the card as expected but on the RTX2070 system this exceeds the cards available memory and the system allocates maybe 3GB from system ram. I wound down all settings to minimum and reduced the rendering to just 30 (which is something like 950x950!!) and with all these settings the GPU memory required is still about 10GB or so of which it allocated 2 to 3GB from system ram. System ram is of course super slow compared to the GDDR6 memory on the card and the performance drops through the floor and is pretty much unusable.

So my question to people is, is this what others are finding and if not, how come. I see many places where people are happily using an RTX2070 card in VR and yet I cannot get the requirement down to below 8GB to match the card. Is this a bug whereby some part of the system or MSFS2020 should recognise the card only has 8GB (even though the OS is saying it could take another 16GB if wanted) which would in some way constrain the data sent to the card.

I have read that the allocating of system memory to graphics is a bios thing although I see no way to change this. It appears not to be a Win 10 thing. Maybe other people have motherboards that do not do this.

Have I got this all wrong? Comments very welcome.

System 1
Core i7-8700K
RTX 2070 Super 8GB
Asus ROG Z370-E motherboard
32GB Corsair 3300MHz Ram
Samsung 960 Pro 512GB NVMe SSD
Samsung 970 EVO 500GB NVMe SSD
HP Reverb G2
Samsung 2160x1440 144Hz monitors

System 2
Core i7-9700K
RTX 3090 FE
Asus Prime Z390-A motherboard
32GB Corsair 3000MHz Ram
Samsung 970 EVO 1TB NVMe SSD
HP Reverb G2
Samsung 2160x1440 144Hz monitors

1 Like
2

I believe you do :slight_smile:

I didn’t read the article but when using any 3D API, you’ll setup buffers for vertices for example. The API (say DX11) lets you build the buffers, therefore it allocates RAM to store the data, prior sending the data to VRAM. The API won’t discard the data though because if the VRAM buffer got discarded, it can re-stream the data back later to VRAM.

3

Thanks for the response.
My understanding of the inner workings of DX11 is very limited to say the least but i think you are saying here that a number of buffers are always setup in normal ram prior to loading up the VRAM and that they may be maintained for later use. What i don’t understand is how, via the API, you would chose to use the system ram part of the available graphics memory for just these buffers and yet use the rest of the VRAM for time critical high speed requirements, doesn’t the entire VRAM appear as just a single blob of graphics memory to the caller? Other posters have said they even sometimes see more than 20GB of memory used on a (3090) card so if that amount was placed onto the 2070, you’re suggesting there would never be more than 8GB that needs to be on the high performance GDDR6?
I clearly need to read up a bit on how a graphics card is driven to fully understand all this but if significant proportions of the used graphics memory use system ram, at what point is performance impacted?
Cheers

4

you might want to start with this:
https://www.qwant.com/?q=dx11+memory+managment

5

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.