NVIDIA RTX 5090 Loses Over 25% Performance Without Full PCIe Bandwidth, With Noticeable Losses in Rendering Workloads

NVIDIA’s flagship Blackwell GPU apparently loses a massive chunk of performance if it isn’t operated under the full PCIe bandwidth, in particular with video editing applications.

Using Any Other Add-In Card Apart From NVIDIA’s RTX 5090 Could Cost a Lot of Performance Difference

Many factors determine a GPU’s performance across multiple workloads, and one important one is the PCIe bandwidth that is accessible to it. With the latest PCIe 5.0 generation, insufficient PCIe lanes could severely deteriorate GPU performance, especially across intensive workloads like video rendering and content creation. Puget Systems has conducted extensive testing on NVIDIA’s GeForce RTX 5090 to determine the performance impact with lesser PCIe bandwidth, and based on the result, it is evident that the difference is quite significant.

Interestingly, PCIe bandwidth allocated to GPUs gets affected if another device is put in a PCIe slot apart from the primary ones, since it essentially distributes the lanes equally. Many motherboard manufacturers have a single PCIe 5.0 x 16 slot, which allows the RTX 5090 to operate at full performance, but if other add-in cards, such as a PCIe network card, are connected, it does affect GPU performance. Puget Systems tested NVIDIA’s flagship Blackwell GPU in rendering applications and AI workloads, and it seemed like the GPU ran perfectly only under full PCIe bandwidth consumption.

Starting with After Effects, the RTX 5090 saw a significant performance hit when it dropped from PCIe 5.0 x 16 to PCIe 3.0 x 4, marking more than a 10% difference. Similarly, with DaVinci Resolve, the performance hit was more than 20% when the GPU ran under PCIe 3.0 x 4, and there were noticeable hits as well when the number of lanes dropped from x16 to x4 across the same generation. This shows that operating multiple add-in cards on your motherboard could affect GPU performance massively.

In Game Dev benchmarks, particularly on Unreal Engine, almost all PCIe configurations showed little change in performance. This was also true for AI workloads, such as Llama.cpp benchmark, where performance was unaffected when lower PCIe bandwidth was assigned to the onboard GPU. The reason why the drop isn’t much significant here is that these applications are more dependent on the GPU VRAM. While the average consumer shouldn’t worry much about PCIe bandwidth, professionals, especially content creators, should keep it in mind.