NVIDIA RTX 5090 Performance Drops 25% Without Full PCIe Bandwidth, Impacting Rendering Workloads

NVIDIA’s leading Blackwell GPU has been found to experience a significant decline in performance when it operates outside of optimal PCIe bandwidth, particularly in applications involving video editing.

Performance Discrepancies When Using Non-NVIDIA Add-In Cards

GPU performance is influenced by several key factors, one of the most crucial being the PCIe bandwidth available. With the rise of PCIe 5.0 technology, inadequate PCIe lanes can lead to substantial reductions in GPU performance during demanding tasks, such as video rendering and content creation. Recent tests by Puget Systems have highlighted the performance ramifications of using lower PCIe bandwidth with NVIDIA’s GeForce RTX 5090, revealing notable differences.

Notably, the amount of PCIe bandwidth available to a GPU can be affected by the installation of additional devices in other PCIe slots, as this can lead to lane sharing. Many motherboards feature a single PCIe 5.0 x16 slot that allows the RTX 5090 to perform at its full potential. However, adding other cards, such as a PCIe network adapter, can limit the GPU’s bandwidth. Testing conducted by Puget Systems demonstrated that NVIDIA’s Blackwell GPU excelled only when it had complete access to the PCIe bandwidth.

Performance impact of reduced PCIe bandwidth

In practical application tests, such as Adobe After Effects, a noticeable performance degradation occurred when switching from PCIe 5.0 x16 to PCIe 3.0 x4—over a 10% drop. Furthermore, in DaVinci Resolve, the GPU’s performance reduced by more than 20% under the same PCIe 3.0 x4 conditions. This highlights how utilizing multiple add-in cards can drastically affect the GPU’s performance capabilities.

Benchmark results of GPU performance across PCIe configurations

However, during benchmarks for gaming development, particularly with Unreal Engine, the variations in PCIe configurations did not significantly impact performance. This observation applied to AI workloads as well, such as those tested with Llama.cpp, where performance remained consistent despite lower bandwidth. The reduced impact in these scenarios is attributed to the greater reliance on GPU VRAM for these applications. While average users may not need to concern themselves with PCIe bandwidth limitations, professionals—especially those in content creation—should certainly take these considerations into account.

Source & Images