Nvidia RTX 5090 loses up to 25% performance with limited PCIE bandwidth, Study points out

A NVIDIA GeForce RTX 5090line video card based on architecture Blackwellshowed performance losses greater than 25% when operating with PCIE Band Width Reduced, according to tests conducted by Puget Systems. The analysis focused on professional workloads and revealed that the plate only reaches its maximum potential when installed in a PCIE PCIE 5.0 x16 exclusive slot.

The tests pointed out that the performance of RTX 5090 in Davinci Resolves It suffers large falls as the bus band width is reduced. With the complete configuration of PCIe 5.0 x16the GPU obtained a score of 12,553 points.

When being limited to PCIe 4.0 x4this number fell to 9,356, and with PCIe 3.0 x4performance plummeted to 6,825 points, a reduction of approximately 45% in relation to the ideal configuration.

Another impacted application was the Adobe After Effectswhich presented more than 10% drop in performance when using only PCIe 3.0 x4.

These results show that the band bandwidth can become a important bottleneckespecially in tasks involving constant transfer of data between CPU, RAM and GPU.

AI and GAMES applications are less sensitive

On the other hand, Benchmarks with Language Models (LLMS) how llama.cpp indicated low variation in performance with different PCIE settings.

Even with limited bandwidth, the “throughput” in tokens per second varied less than 10% between tests. In this case, the GPU uses a lot of I do give it a gpuwhich reduces the band’s dependence between the components of the rest of the PC.

In applications like the Unreal Enginethe performance reduction was also small. Most game development tests showed differences within the margin of error, even with PCI limited to X8 or X4.

Use of additional plates can have an impact

One of the main factors that can limit the bandwidth available for the RTX 5090 It is the use of Other expansion boards No system.

In motherboards that share PCIE lanes between slots, installing an additional network or SSD card can force the GPU to operate in X8 Ou X4compromising the performance.

Most current motherboards offer Just a PCIE PCIE 5.0 x16 With total bandwidth, while the others operate in more limited modes.

This is due to the division of Lanes to support multiple devices such as M.2 drives, USB controllers and network interfacesleaving less resources for the GPU.

Recommendation for professionals

To content creatorsvideo studios and users who depend on intensive rendering applications, the recommendation is clear: avoid sharing LANES PCIE with other devices and checking the system configuration using tools such as GPU-Z.

If necessary use Multiple GPUSmake sure the motherboard keeps at least PCIe x8 by gpu. In legacy systems with PCIe 3.0impact may vary between 4% and 25%, depending on the application.

Without these precautions, you may be taking advantage of only a portion of the total performance that your (extremely) face card can offer your workloads.

That is, professionals must be aware of the Slot Settingshas compatibility with PCIE 5.0and to simultaneous use of other plates. Ensuring that the GPU operates with total bandwidth is essential to justify the investment in a high performance video card like RTX 5090.

Source: Puget Systems. Via: TechPowerUp.