RTX 5090 PCIe 5.0 vs PCIe 4.0: Are You Losing Performance on an Older Motherboard?

For gaming, barely. TechPowerUp measured an RTX 5090 about 1% slower on PCIe 4.0 x16 than on PCIe 5.0 x16, at every resolution. Gamers Nexus found at most about 4% even on PCIe 3.0. If you're gaming on a good PCIe 4.0 board, the slot isn't a reason to replace it; your CPU matters far more. The picture changes for some creative work: Puget Systems saw DaVinci Resolve lose about 10% at PCIe 3.0 x16-class bandwidth and about 25% at lower link widths. It also changes when you push AI models out of VRAM or split them across several GPUs. Narrow chipset slots (x4) cost more everywhere.

How much bandwidth does each PCIe link give?

Source: TechPowerUp RTX 5090 PCI-Express scaling review (Jan 2025); bandwidth calculated from PCI-SIG line rates with 128b/130b encoding. Gamers Nexus (13 Feb 2025) found Gen 4 about equal to Gen 5, with Gen 3 at most about 4% behind (Black Myth: Wukong at 1080p).

Why so little? Once a game's textures and geometry are in the card's 32 GB of VRAM, the GPU reads them at 1,792 GB/s. That's about 28 times faster than even a PCIe 5.0 x16 link. The bus is mostly used for streaming new assets and sending draw commands, which needs far less bandwidth than the card has. TechPowerUp saw no measurable loss in ray tracing at PCIe 4.0 x16, with upscaling slightly more sensitive than native rendering.

Where does PCIe 4.0 actually cost performance?

Puget Systems (3 Jul 2025, Core Ultra 9 285K, ASUS ProArt Z890-Creator) tested content-creation apps:

• DaVinci Resolve was the most sensitive. PCIe 5.0 x16, 5.0 x8 and 4.0 x16 performed about the same. At 3.0 x16-class bandwidth it ran at about 90%. At 3.0 x8 or 4.0 x4 it fell to about 75%, and at 3.0 x4 to about 54%.
• Unreal Engine 5.5 held about 93% at 4.0 x4 or 3.0 x8.
• After Effects was within margin down to 16 GB/s.
• Blender and Octane showed no meaningful effect once the scene was loaded.

So a 5090 on PCIe 4.0 x16 loses essentially nothing in these apps. The losses come from narrow links: x8 on an older Gen 3 or Gen 4 board, or a chipset x4 slot.

PCIe 4.0 x16 is fine for an RTX 5090. A narrow x4 or Gen 3 x8 link is where the real losses start.

What about local AI?

For a model that fits in 32 GB of VRAM, the link mostly affects load time. Puget's small llama.cpp test (Phi-3 Mini) showed no consistent pattern across link speeds. Puget also cautioned that multi-GPU setups and models partly offloaded to system RAM "could see a large effect". That follows from the numbers. When a model spills out of VRAM, the GPU fetches weights across PCIe, and 31.5 GB/s instead of 63 GB/s halves that path's bandwidth.

Multi-GPU work leans harder on the bus. GeForce cards have no NVLink, and NVIDIA says peer-to-peer transfers between GeForce 50-series cards aren't supported. Traffic between two 5090s goes over PCIe via system memory, so slot width matters more there than for one card. We found no controlled PCIe study for multi-5090 setups. See two RTX 5080s vs one RTX 5090 and why 32 GB still limits AI developers.

Could your motherboard be running the card at x8 without you knowing?

Quite possibly. Many boards share PCIe lanes between the main GPU slot and M.2 slots. On the ASUS ROG Crosshair X870E Hero, for example, ASUS states that enabling the M.2_2 slot drops the top x16 slot to x8. On a PCIe 5.0 board that's 5.0 x8, the same bandwidth as 4.0 x16, which costs about 1% in games. On a PCIe 4.0 board the same split gives 4.0 x8, equivalent to 3.0 x16, costing about 2–4% in games and more in Resolve.

Check your motherboard manual's lane-sharing table before adding drives, and check the link actually negotiated using a tool such as GPU-Z. Never put a 5090 in a chipset x4 slot if a full x16 slot is available.

What about cards with only eight lanes?

The RTX 5090 uses all 16 lanes, but the RTX 5060 Ti uses only eight, so the board's PCIe generation matters more for it. TechPowerUp found the 5060 Ti 16 GB about 2% slower on PCIe 4.0 and about 4% slower on PCIe 3.0. ComputerBase found the 8 GB version averaging about 6% slower on PCIe 4.0 than 5.0 at 1440p across 27 games, with some titles losing much more. When an 8 GB card runs out of VRAM, assets stream over a link that is both narrower and older. See RTX 5090 vs RTX 5060 Ti 16GB.

Should you upgrade your platform for an RTX 5090?

• PCIe 4.0 board, fast CPU: keep it. The slot costs you about 1%.
• PCIe 3.0 board: expect a few per cent in games and more in video editing. The CPU on such a board is likely the bigger problem.
• Older CPU, any board: this is usually the real bottleneck. With a 5090, TechSpot measured a Ryzen 7 9800X3D 37% faster than a Ryzen 5 7600X at 1080p Medium and 8% faster at 4K Ultra. More in why your RTX 5090 might never reach its full potential.
• Multi-GPU AI or heavy Resolve work: prioritise a board that gives each GPU at least x8 from the CPU at PCIe 5.0.

A PCIe 5.0 platform does have other benefits, such as Gen 5 NVMe storage and newer CPUs, but those are reasons in their own right, not because the GPU needs the slot.

Buying in Singapore

RTX 5090 cards are not in stock at SourceIT at the time of writing; ask us about availability and a quotation. If you're moving to AM5 for a 5090 build, the AMD Ryzen 7 9800X3D is S$756.38 and the Ryzen 9 9950X3D is S$1,108.02 (both listed 6 Nov 2026). We don't list motherboards, so ask us to source a board with the lane layout you need. Prices change, so ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

Does an RTX 5090 work in a PCIe 4.0 slot? Yes. PCIe is backward compatible, and TechPowerUp measured only about a 1% drop in games on PCIe 4.0 x16.

Is PCIe 3.0 too slow for an RTX 5090? In games, it costs about 2–4%. In DaVinci Resolve, Puget measured about 10% lower performance at that bandwidth. Older PCIe 3.0 platforms usually have a CPU bottleneck that matters more.

Does PCIe 5.0 matter for AI? Not much for a single GPU running a model that fits in VRAM. It matters more when models spill into system RAM or when several GPUs share work over PCIe.

Why is my RTX 5090 running at x8? Many motherboards split the top slot's lanes when certain M.2 slots are used. Check your board's manual and the link speed reported by GPU-Z.

Is 5.0 x8 the same as 4.0 x16? Yes, both give about 31.5 GB/s per direction. TechPowerUp measured both at about 99% of a full 5.0 x16 link in games.