PCIe Bottlenecks Slash NVIDIA GeForce RTX 5090 Content Creation Performance by 25%

[AleksandarK]

A lack of PCIe bandwidth can reduce RTX 5090 performance by as much as 25% in video editing and 3D workloads. Puget Systems conducted a series of tests on NVIDIA's GeForce RTX 5090, comparing different PCIe versions and lane configurations to assess the impact on content creation tools. At its launch, TechPowerUp's scaling tests noted that the GeForce RTX 5090 is the first high-end card to adopt PCIe 5.0 x16, delivering 64 GB/s of bidirectional bandwidth—twice the 32 GB/s that the RTX 4090 achieves on PCIe 4.0 x16. Our testing revealed a potential performance impact of up to 25% when utilizing limited PCIe bandwidth. Puget Systems today put this interface to the test, confirming real‑world impacts on video editing workflows when lanes or generations are constrained.

In DaVinci Resolve benchmarks, Puget Systems found that configurations running at PCIe 5.0 x16, PCIe 5.0 x8, or PCIe 4.0 x16 yielded virtually identical render times. Dropping to PCIe 5.0 x4, PCIe 4.0 x8, or PCIe 3.0 x16 introduced a modest 10% slowdown. Further reducing bandwidth to PCIe 4.0 x4 or PCIe 3.0 x8 resulted in an increase of roughly 25% in render times. After Effects exhibited only minor slowdowns once the bandwidth dropped below 8 GB/s. Unreal Engine 5.5 virtual production tests recorded about a 7% drop in average frame rates at the lowest lane counts. By contrast, Blender offline renders and OctaneBench scores remained essentially unchanged, and Llama LLM benchmarks showed no measurable dependency on PCIe speed.




The PCIe 5.0 on RTX 5090 employs the same NRZ signaling as PCIe 4.0 but incorporates stricter signal integrity measures, including decision feedback equalization and tighter timing controls. Although fully backward compatible with PCIe 4.0, 3.0, and even older standards, the RTX 5090's massive bandwidth headroom raises the question of how much performance is lost on legacy slots or when lanes are shared with NVMe drives. These findings reveal a common limitation of modern motherboards: most reserve full x16 lanes for a single slot, forcing any additional devices or drives to operate at half or quarter speed. As a result, a flagship GPU can unknowingly operate at PCIe 4.0 x4, extending project turnaround times for professionals working with high-resolution timelines or complex 3D scenes. For studios and power users planning multi-card configurations, the clear recommendation is to verify that the RTX 5090 remains in a full-bandwidth slot.

View at TechPowerUp Main Site | Source

 

[Daven]

Lower PCIe lane counts should only be used for I/O expansion cards. All GPUs should come with PCIe 5 x16 and greater.

 

[agent_x007]

"s" We here like clickbait ! "/s"

Actual title should be like this :
Performance of RTX 5090 drops by up to *25% !
(*when you run it at PCIe 4.0 x4, PCIe 3.0 x8 or PCIe 2.0 x16)

 

[_roman_]

I see it differently.

that news piece shows pcie 4.0 x 16 electrical lanes is still enough. Overspending on a pcie 5.0 mainboard may not be wise when building on a restricted budget.

In my point of view pcie 5.0 is a hoax. pcie 5.0 NVME are barely faster as my KC3000 2TB @ pcie 4.0. It seems this is the same for graphic cards.

AMD B450, afaik am4 chipset, mainboards or similar dated hardware slowly get issues with new graphic cards. AFAIK PCIE 3.0 on the graphic card slot.

 

[londiste]

I will just leave this summary bit from Puget here:

On modern motherboards, you often only get one PCIe slot at a full 5.0 x16 bandwidth. Additional slots may be 5.0 x8, but are likely much lower, at 4.0 x4 or below. Because of this, multi-GPU setups or configurations with add-in cards may find one or more GPUs with dramatically reduced PCIe bandwidth. Although most of the workflows we tested don’t show too much performance loss at 4.0 x4, that’s not true across the board.

In video editing/motion graphics, we saw the largest impact. PCIe 5.0 x16, x8, and 4.0 x16 were functionally equivalent. However, below that, we started to see some differences, especially in DaVinci Resolve. In that application, 3.0 x16 was 10% slower, and our typical-case 4.0 x4 was about 25% slower. These margins are reduced in After Effects, but still present. We recommend caution when configuring a system for video editing applications with multiple add-in cards, as reducing the number of lanes available to the GPU can have a measurable impact on performance.

Our Unreal Engine benchmark also showed performance impacts from PCIe bandwidth. However, the impacts are more minor. We only saw a noticeable hit once the bandwidth was reduced to 4.0 x4 (or equivalent), with an average fps drop of 7%. 3.0 x4 was slightly worse, at 10% slower than maximum bandwidth. While we are less concerned about this amount of lost performance, it should still be kept in mind.

Offline renderers and LLM benchmarks showed no impact from PCIe bandwidth on performance. This makes sense as both tend to load their work fully into GPU VRAM and crash if they can’t. There are some exceptions to this with LLMs, but operating out of system RAM is a huge slowdown. Thus, while reduced PCIe bandwidth may slow initial model or scene loading, it should have a negligible impact on performance after that. Our one note of caution here is that, in situations where you are pooling VRAM to fit a model, PCIe bandwidth may have a large effect. We were not able to test that here.

When we configure the systems we sell, we balance the need for maximum performance from components with the desire for add-in cards necessary for our customers to do their work. Frequently, this means reducing the primary GPU to PCI-e 5.0 x8, which reduces the PCI-e bandwidth in half. However, as we showed in this article, this major reduction in bandwidth often has a minimal impact on real-world performance. Outside of a few uncommon situations, this testing confirms that as long as you have a modern motherboard that supports PCIe 5.0, running the GPU at x8 speeds is not an issue. However, lower-end motherboards, which will require the GPU to run at 4.0 x4, may introduce performance penalties.

 

[dgianstefani]

I see it differently.

that news piece shows pcie 4.0 x 16 electrical lanes is still enough. Overspending on a pcie 5.0 mainboard may not be wise when building on a restricted budget.

In my point of view pcie 5.0 is a hoax. pcie 5.0 NVME are barely faster as my KC3000 2TB @ pcie 4.0. It seems this is the same for graphic cards.

AMD B450, afaik am4 chipset, mainboards or similar dated hardware slowly get issues with new graphic cards. AFAIK PCIE 3.0 on the graphic card slot.

People buying 5090s often on tight budgets do you think?

 

[freeagent]

I am impressed, 5090s have dropped from 6500CAD to under 4500CAD. except for the Asstrails, those are still hanging out near the 5K mark

 

[TheinsanegamerN]

People buying 5090s often on tight budgets do you think?

Having a larger budget doesnt mean you NEED to splurge on things you dont need.

Besides, I can guarantee you there are lots of 5090 buyers who went into debt to buy the thing. Those types of luxury items tend to draw in people who really cant afford them in the first place (see also: nice cars, expensive houses, "luxury" vacation packages, "luxury" clothing, ece).

I see it differently.

that news piece shows pcie 4.0 x 16 electrical lanes is still enough. Overspending on a pcie 5.0 mainboard may not be wise when building on a restricted budget.

In my point of view pcie 5.0 is a hoax. pcie 5.0 NVME are barely faster as my KC3000 2TB @ pcie 4.0. It seems this is the same for graphic cards.

AMD B450, afaik am4 chipset, mainboards or similar dated hardware slowly get issues with new graphic cards. AFAIK PCIE 3.0 on the graphic card slot.

Before PCIe, the bottleneck on AM4 will be the CPU. Sure they're fine, but there's a reason they use the 9800x3d to test games on the 5090. On AM4, your limit is the 5800x3d, which is roughly the same as a 7800. That will become a "restriction" before the PCIe bus will.

PCIe is a far worse issue on low VRAM cards, like the 8GB abominations we've been getting, once stuff has to be swapped to system RAM performance tanks. Even 5.0x4 is vulnerable to this.

 

[AnotherReader]

I am impressed, 5090s have dropped from 6500CAD to under 4500CAD. except for the Asstrails, those are still hanging out near the 5K mark

There are some SKUs that are around the 3500 CAD mark with one going for 3300.

 

[Tyler-98-W68]

There are some SKUs that are around the 3500 CAD mark with one going for 3300.

View attachment 406563

I got my Tuf 5090 PC for 3629 cad$

The FEs are 2899 but impossible to get (best buy exclusive) seems most things in canada are now available at MSRP. I wanted a FE but settled for a TUF OC, didn't want the gigabyte cards for their leakage.

And here my OCD was going wild think 5.0 8x would make a difference when I install a second gen 5 ssd. Good to know it wont

 

[freeagent]

There are some SKUs that are around the 3500 CAD mark with one going for 3300.

View attachment 406563

I was just looking at Amazon.. I keep forgetting there is a world outside

 

[thesmokingman]

DUH?

 

[freeagent]

Cheaper than Newegg surprisingly..

 

[Panther_Seraphin]

Interesting to see something that shows meaningful drop from going from PCIe 3/4 as gaming I believe is still single digit figures all the way down to 2.0 16x equivalents

 

[TheinsanegamerN]

Interesting to see something that shows meaningful drop from going from PCIe 3/4 as gaming I believe is still single digit figures all the way down to 2.0 16x equivalents

The beauty of games, as opposed to pro applications, is that almost all the info the GPU needs can fit in their VRAM pools. Once that initial loading is done bandwidth really isnt an issue anymore.

That's why 4GB, and now 8GB, cards show significant performance regression on cut down busses compared to larger VRAM cards, because they have to swap out stuff with system RAM and saturate that PCIe bus whenever they do so.

 

[Athlonite]

What I'd like to know is what idiot would be buying a 5090 only to run it on a PCIe 3.0 mobo even at x16.
It would seem that this is little more than a make work test done by someone with nothing better to do but who doesn't want to pick up a broom and sweep the floor

 

[Tyler-98-W68]

What I'd like to know is what idiot would be buying a 5090 only to run it on a PCIe 3.0 mobo even at x16.
It would seem that this is little more than a make work test done by someone with nothing better to do but who doesn't want to pick up a broom and sweep the floor

This idiot....because I wanted to see how it worked....and it was just fine.

 

[iameatingjam]

People buying 5090s often on tight budgets do you think?

How tight are we talking? I bought a 4090, and for while there people seemed to think that meant I had unlimited money for other components and I didn't.

I mean it was like, you buy a xx90 class card and people act like its a sports car or something. I understand the 5090 is more expensive still but I don't think the disconnect has changed. Just because you bought a 5090, doesn't mean you have to go buying motherboards with features you don't need.

There's people who can just barely scrape up the money to buy one, and there's people where its nothing but a rounding error on their bank account. Not the same thing.

My hobby before I got back into computers, ancient coin collecting was totally ruining me financially (well kinda, I mean I still have them, so I could sell them if I wanted, potentially even for more than I bought them for, but I don't wanna....), anyway I got back into PCs because I needed something cheaper to occupy my time. And cheaper it certainly is.

PC building is cheaper than a lot of hobbies, even taking the price of the 5090 into account.

 

[blahoink01]

People buying 5090s often on tight budgets do you think?

You aren't wrong, but there is a use case for 4.0. When I built with the 5090, I vertically mounted it. The only compatible vertical mount for my case at that time was a 4.0 expansion cable with its raised holder. So the slot is 5.0 that I have to manually set to 4.0 to make it work with the riser.

 

[dgianstefani]

You aren't wrong, but there is a use case for 4.0. When I built with the 5090, I vertically mounted it. The only compatible vertical mount for my case at that time was a 4.0 expansion cable with its raised holder. So the slot is 5.0 that I have to manually set to 4.0 to make it work with the riser.

There's 5.0 certified risers now, have been for about a year I think, about $70-100 from memory.

 

[_roman_]

People buying 5090s often on tight budgets do you think?

A vallid argument. So all those guys who wrote on those german forums are wrong, right?

The usual upgrade question was: Is it a CPU problem (limit) or a GPU problem (Iimit).

Just because a nvidia 5090 was used to determine the influence of the bus speed does not change the facts about the bus speed.

My comment was purely on the bus speed of the combination of a mainboards with a cpu.

--

Do all tasks for a graphic card needs a ryzen 9800x3d or better processor with pcie 5.0 mainboard? Yes? I'm sorry i was so wrong.
e.g. I think those mining mainboards only had a single pice lanes each. I never was in teh gpu mining thing.

 

[mashie]

I see it differently.

that news piece shows pcie 4.0 x 16 electrical lanes is still enough. Overspending on a pcie 5.0 mainboard may not be wise when building on a restricted budget.

In my point of view pcie 5.0 is a hoax. pcie 5.0 NVME are barely faster as my KC3000 2TB @ pcie 4.0. It seems this is the same for graphic cards.

AMD B450, afaik am4 chipset, mainboards or similar dated hardware slowly get issues with new graphic cards. AFAIK PCIE 3.0 on the graphic card slot.

Building to a restricted budget with a 2k+ 5090?

 

[lexluthermiester]

What I found interesting from those results is that PCIe 4 x16 still performs perfectly and PCIe 3 x16 still performs very well. Even PCIe 2 x16 performs acceptably.

 

[iameatingjam]

I'm happy to know pcie 4x16 is still doing well on the high end. Although, this is a best case scenario. As soon as you start needing to transfer dram through the bus, or there's only 8 lanes on the card... or both... then things might be different. What if said pc only has ddr4? Then it may even be worse.

Also.... this headline is kinda clickbaity.

 

[Bomby569]

lets try the worst scenario and create clickbait, not very different from the million videos on the 8gb vram clickbait from HWU

this type of crap is fun when someone puts a 5090 on a 10 year old optiplex for the laughs.