PCI express bandwidth with two GPUs is a specific configuration.
PCI express bandwidth with two GPUs is a specific configuration.
Hello, I own a 5800X3d CPU with 32GB RAM and an ASUS B550-E motherboard. My graphics card is an RTX 3080. I’m considering adding a second GPU into the second PCIe slot. Both cards would need to run at x8 speeds. I’m thinking about purchasing a budget GTX 1060 3GB for the second card, as they are very affordable. The RTX 3080 uses PCIe 4, while the GTX 1060 uses PCIe 3. The GTX 1060 3GB will operate with PCIe x8 version 3 bandwidth, which is acceptable. Regarding the RTX 3080, if it reaches PCIe x8 version 4 speeds, that’s fine. If it drops to PCIe x8 version 3, then there could be issues. What speed would the RTX 3080 achieve in this setup? I’m aiming for CUDA-based applications. For gaming, a dedicated GPU is ideal for CUDA workloads and fluid simulations—performance can drop with a maxed-out GPU. A dedicated frame generation card using lossless scaling would help avoid stuttering. It would be a much cheaper alternative compared to buying RTX 4000/5000 cards. Any helpful advice would be appreciated. Thanks.
The task is clear—provide a revised version of what was asked.
The performance should actually decline when modifying the frame buffer content directly on the main card. Transferring data between cards isn't efficient. It's unclear if the games support offloading computation to a different GPU for rendering.
Perhaps it doesn't address my question, but I'll look into the lossless scaling forums for more details. I'm sharing this to find info on PCIE bandwidth problems. Even if it's useful for the Killing Floor 2 game, it's still worth purchasing for me.
No we understand what you got confused on, physX (cuda) on a seperate card is a thing, but its not anymore and hasnt been for about 10 years. And using a slower GPU for computing physX actually slows down your game. notice the GTX 580 and the 8600GTS physX card results. My comment is on the video still I never did go back to having a dedicated physX card because a GTX 760 would slow down my RTX 3060 as shown by the chart using GT 730as physX cards. and for all of a couple of games? a GTX 1060 will compute the physX SLOWER then the RTX 3080 can compute the render and the physX. meaning the 1060 will be your bottleneck. I also dont get free electricity anymore so I dont keep the second GPU in the PC to inefficiently fold@home either so its not even there to mess with. If you want to dedicate a GPU to frame generation, you need one with Tensor cores, a GTX 1060 will not work for your wants. Anyways, if you go forward with the experiment, just try it, 50% chance it works the way you want it to
The GTX 1060 is best suited for frame generation only, while another card should handle rendering. I recommend learning about the lossless scaling tool. Performance depends on your settings and the game. I mean by flex, not PhysX. If you use PhysX, just pair the same card for both tasks. What I wrote above isn’t directly related to what you’re asking. The motherboard layout and table are helpful. My ASUS B550-E board has a setup that matches yours. Both GPUs connect to the CPU via a chip that splits lanes. I’m curious: if slot 1 has a PCIe 16x version 4 (marked 1) and slot 2 has a PCIe 16x version 3 (marked 3), will the first card run at PCIe x8 version 4 speed (~16GB/s), or will the first card slow down to ~8GB/s? If the first card’s speed drops, that could be an issue. I tried to explain clearly as best I could! Thanks.
Bottom slot is limited to x4 configuration and cannot be upgraded further. Slots 1 and 2 will operate in x8 mode together. I’m unsure about the PCIe version they’ll use—likely different from each other, though it’s unclear. This diagram represents your specific motherboard layout, not a generic one.