Not generally, but there are a few edge-cases. AMD recently released the 6500XT GPU which has PCIe 4.0 but only x4, and when used in PCIe 3.0 motherboards that

Not generally, but there are a few edge-cases. AMD recently released the 6500XT GPU which has PCIe 4.0 but only x4, and when used in PCIe 3.0 motherboards that obviously becomes 3.0x4. Furthermore, AMD only put 4GB of VRAM on it, which isn't good enough for modern games (it's actually less than the previous 5500XT, where AMD made a big advertising push about "4GB isn't enough for modern games anymore") so it swaps all the time, which amplifies the PCIe bottleneck. The card loses something like 15% performance when used on a PCIe 3.0 motherboard. It's not like it won't run, but that's a significant amount of performance, that's like a half a tier of performance and far more than you see on (eg) 3090. Unfortunately we're still in something of a transition period. AMD blocked vendors from supporting PCIe 4.0 on 300/400 series boards that might have been capable of the required signal integrity (particularly on the first slot). AMD doesn't support PCIe 4.0 on their Zen3 APUs at all either - and some of the new processors are based on the APU die even though they don't have graphics, so they are limited to PCIe 3.0 as well. And obviously Skylake/Coffee Lake/Comet Lake stuff is all PCIe 3.0 based since they're ancient. So there are definitely scenarios where you might think "throw in a cheap dGPU" and are still stuck with PCIe 3.0. Anyway though, what I would caution you here is, the 3090 has lots of VRAM, so it doesn't swap. What the 6500XT shows is, low-end cards can be more susceptible to PCIe bottlenecking - they have less VRAM, so they swap more, which increases the pressure on the PCIe bus. 3090 results are not representative of a worst case scenario just because they do drawcalls really fast (high framerate), there are other parts of the pipeline where PCIe load can be generated. If you are swapping due to low VRAM, that's still PCIe load. Similarly - simply using a card placed into a PCIe 3.0x4 slot is not representative of Thunderbolt 3 results either, despite both links being 3.0x4. Thunderbolt is usually implemented as an additional standalone chip attached to the chipset - so it's not a CPU-direct link, there's multiple hops with higher latency there. There is also contention for the chipset bandwidth - the chipset has a 3.0x4 link, so the GPU alone can saturate it, but there's also NVMe traffic (particularly for pre-11th gen Intel where NVMe has to attach to the chipset) and network traffic (chipset provides the network controller), etc. It's bidirectional so you can read at 4GB/s while you write 4GB/s to the GPU, but there's also just general contention for commands/etc. So performance results on Thunderbolt will be worse than a card attached to the chipset, which will be worse than a card attached to CPU-direct PEG lanes, even if lane count is the same. (the exception to this might be Ice Lake and newer Intel laptop chips, where the Thunderbolt controller is actually part of the CPU itself, the performance impact of that should be less. However, this does not apply to desktop chips, including Alder Lake.)