NVIDIA GeForce RTX 5090 PCI-Express Scaling

[Cadasiso]

I was planning on building a new AM5 system along with a 5080 but might just keep my Taichi x370 PCIe 3.0 mobo and get a 5090 with the savings of not buying a new mobo and RAM. I play at 4k. What do you guys think?

It'd be fine, just slap a 5700X3X there and you're good for Ultra+RT, I don't think bottleneck is a major concern in 4K in newer, GPU bound games, it's not like there's any 4090 for reasonable prices anywhere in the world anyways, so even while bottlenecked it'll be a better deal (assuming you can get one). CPU scaling reviews might come soon, it'd wait for them before considering between the 5090 and 5080

 

[W1zzard]

Why didn't you compare the stutter intensities when VRAM overflows?

Still waiting for that 32 GB VRAM game

 

[_roman_]

It'd be fine, just slap a 5700X3X there and you're good for Ultra+RT

No - definitely not - a recent retest from a gaming magazine

Proof: https://www.pcgameshardware.de/Benc...iste-Index-Gaming-2025-Fps-Vergleich-1464279/

well Numbers are a universal metric which should be understood by mostly anyone.

The 5700X3d will cost you a lot of frames because of the outdated, old plattform. Ultra and Raytracing implies high end. Not entry level gaming, like gamig wiht a low cheap, outdated, old, ryzen 5700X3d with old DRAM Standard, old PCIE standard and such. that processor is also not really future proof for ultra and raytracing, e.g. year 2025, year 2026, year 2027

Ultra is above very high - is above high - is above medium high - is above light medium - is above medium - is above low - is above low.

Just to make clear - we are talking about ULTRA game settings with Raytracing. Not low end

 

[Haunter]

I really didn't expect that. Once I was testing different Pcie and Procyon, AI workload including some Time Spy and games, etc. are in average utilizing PCIE 4 on about 10%. AI workloads more, almost full bandwidth. So this comparison is really nice and detailed. And I really didn't expect 4080 to be better, than best card in the world, LOL, great step back, nVidia, really great... first nvlink, now this. Same like car manufacturers using eco leather...

 

[trsttte]

Too busy, 5090 custom, then more releases from NV, and maybe AMD manages to launch something, too

I get it, but if you could squeeze just a few tests even without comparing to previous generations would be worthwhile, this is something a lot of people will be looking for

How to test lower PCIEX generations with same motherboard?

Bios settings to either set the pcie generation or limit the number of lanes. If the motherboard doesn't have at least one of these options duct tape on the gpu pcie contacts does the trick.

Funny thing is can see Nvidia releasing it with only 4 lanes.

Probably not x4, but likely x8 like the 4060

 

[lexluthermiester]

Still waiting for that 32 GB VRAM game

That makes me wonder: What would a 5090 do at 8k?

 

[mouacyk]

damn i need my 1%

 

[nageme]

So to quickly sum up, PCIe bus versions are not a limiting factor to GPU performance.

A game or two there did show a more pronounced slowdown, at least at 1920x1080. I'm guessing it depends on how the game is designed.
Maybe games have learned, historically, to avoid relying too much PCIe traffic because of bandwidth limitations (or latency?). But as PCIe speeds up, new ideas become possible.

For example (though not really related to graphics bandwidth), old games never streamed level data. Maps always pre-loaded fully at the beginning, or at best at a few checkpoints.
But at some point someone realized it could be done. Later, others followed.

Most motherboards should have the option to limit the speed of PCIe in the bios/uefi.

What's it meant to be used for?

 

[lexluthermiester]

A game or two there did show a more pronounced slowdown, at least at 1920x1080. I'm guessing it depends on how the game is designed.

True, very situational. However, none of those results were anywhere near unplayable until you get below PCIe 2.0. 3.0 and 4.0 are still perfectly acceptable. With most titles one wouldn't notice the difference at all.

 

[Slackaveli]

I would love to see motherboards that only provide 5.0 x8 (essentially 4.0 x16) bandwidth to the GPU and reallocate the other 5.0 x8 lanes to M.2 and/or additional PCIe slots. I'd rather Give up 1-2% performance for much more expandability.

then you need my board- 5 gen5 at the same time on Godlike x870e.

 

[lexluthermiester]

Jay did a set of runs comparing and his numbers mimic the numbers W1zzard got;

So in reality, PCIe gen is not a factor for performance.

 

[Mysteoa]

Hey, @W1zzard. Can you run some PCIE-ex Test with RX 9070XT, If you have the time? Nothing extensive, just few benchmarks to check if it works as expected.
I have stumble upon a few people that are claiming there is a difference between PCI-Express 5, 4 and 3 on AMD. Logically, it should be the same as the 5090. I would have tested my self, but I don't have access to PCIEx5.

 

[blinnbanir]

This is how it is already done. Take AMD's AM4 and AM5 platforms for example. These have 24 PCIe lanes, divided as 4 lanes to motherboard, 4 lanes to NVME, and 16 lanes to PCIe. Most motherboards will have more than a single NVME slot. One will use the dedicated x4 lanes, the others will "steal" lanes from the PCIe, or piggyback on the Mobo lanes.

That is not true

 

[_roman_]

#63

read the mainboard manual first how the slots and lane sharing is. It can be true for certain hardware configurations.
It also depends on teh processor in question, which is also explained in the mainboard manual. This also applies for devices which are assembled together like tablets, smartphones and laptops. Anything which has some sort of main logic board or named mainboard.

 

[blinnbanir]

#63

read the mainboard manual first how the slots and lane sharing is. It can be true for certain hardware configurations.
It also depends on teh processor in question, which is also explained in the mainboard manual. This also applies for devices which are assembled together like tablets, smartphones and laptops. Anything which has some sort of main logic board or named mainboard.

I have a X870E Taichi and guess how many PCIe lanes it has? 44 lanes. That is right there is enough PCIe lanes to have no lane sharing. I only have 1 SSD port left on my MB but I could use the USB 4.0 x2 ports with USB C adapters to add more way more storage. That is 30TB of storage there no handicap for PCIe lanes for X870E my X670E board had even more M2 slots so the same hold true for some X670E boards. Remember they come with 2 chipsets.

BTW there are 3 RAID 0 arrays in my system

 

[tugrul_LordOfDrinks]

My ssd corrupt after a failed cold boot after a failed gpu oc attempt.

Is ssd dying or was it oc fault?

Windows repaired itself and survived but some games couldnt.

 

[trsttte]

I have a X870E Taichi and guess how many PCIe lanes it has? 44 lanes.

AM5 only has 28 lanes, where does the motherboard get the remaining 16 from? It doesn't!

There is lane sharing, the 2 chipsets handle it, turning 4 lanes into 12 to allow for 3 extra m.2 drive (I dunno how you arrived at the 44 number, whatever way you used is wrong). The block diagram for the board shows this very clearly (kudos to Asrock for still making them easily available in the manual)

Btw, not only there's lane sharing, they're sharing the bandwidth with everything else in the system - USB, wifi, SATA and the sound DAC. You'll only get to use the SSD full speed one at a time and only if you're not doing much else on the computer. I personally don't understand why people bother with this kind of set up but hey you do you.

 

[blinnbanir]

AM5 only has 28 lanes, where does the motherboard get the remaining 16 from? It doesn't!

There is lane sharing, the 2 chipsets handle it, turning 4 lanes into 12 to allow for 3 extra m.2 drive (I dunno how you arrived at the 44 number, whatever way you used is wrong). The block diagram for the board shows this very clearly (kudos to Asrock for still making them easily available in the manual)
View attachment 393897

Btw, not only there's lane sharing, they're sharing the bandwidth with everything else in the system - USB, wifi, SATA and the sound DAC. You'll only get to use the SSD full speed one at a time and only if you're not doing much else on the computer. I personally don't understand why people bother with this kind of set up but hey you do you.

Are you suggesting that I cannot use 2 drives at the same time? Are saying that I don't have access to my M2 drives while looking at my SATA drives. I love that notion, Now you are saying that my speed will be cut in half. When I say lane sharing I am talking about if you plug a M2 into M2_4 for example you would lose access to 2 SATA ports. Now I also see that you are suggesting that if I am listening to music I am losing performance on my SSDs too? BTW what is the theoretical bandwidth of PCIe 4.0x4. 8GB/s. Let me see? So I am listening to Music. While doing that I ran a benchmark one 1 drive connected to the Chipset ands downloaded a Game to another drive connected to the Chipset.

https://www.amd.com/en/products/processors/chipsets/am5.html

 

[trsttte]

Are you suggesting that I cannot use 2 drives at the same time? Are saying that I don't have access to my M2 drives while looking at my SATA drives. I love that notion, Now you are saying that my speed will be cut in half. When I say lane sharing I am talking about if you plug a M2 into M2_4 for example you would lose access to 2 SATA ports. Now I also see that you are suggesting that if I am listening to music I am losing performance on my SSDs too? BTW what is the theoretical bandwidth of PCIe 4.0x4. 8GB/s. Let me see? So I am listening to Music. While doing that I ran a benchmark one 1 drive connected to the Chipset ands downloaded a Game to another drive connected to the Chipset. View attachment 393908

https://www.amd.com/en/products/processors/chipsets/am5.html

No, i'm saying they can't all do full speed at the same time, because they can't as you can see on the image I sent and on any AM5 architecture diagram. Listening to music is obviously a rouding error, run crystaldisk on 2 drives at the same time and you'll see the performance drop. That board doesn't have the 2 SATA or 1 m.2 thing that was common with AM4 but it still does lane sharing, the CPU only has 28 after all.

 

[ARF]

Hey, @W1zzard. Can you run some PCIE-ex Test with RX 9070XT, If you have the time? Nothing extensive, just few benchmarks to check if it works as expected.

I have stumble upon a few people that are claiming there is a difference between PCI-Express 5, 4 and 3 on AMD.

Link with benchmarks?

Logically, it should be the same as the 5090. I would have tested my self, but I don't have access to PCIEx5.

The slower the card (RX 9070 XT is slower), the lower the PCIe throughput impact.
PCIe 5.0 is absolutely useless today.
It is a waste of precious lanes which otherwise could have been used for SSDs, CrossFire, etc.

 

[tugrul_LordOfDrinks]

More pcie bandwidth helps when there is destructible terrain and cpu computes the destruction. The updated terrain data can be gigabytes and it takes 20 milliseconds to copy 1 GB on pcie 5.0.

For example in a minecraft clone, you use TNT and blow up half of map. 10 million cubes are destroyed. If 1 cube takes 100 bytes, it updates 1GB. 50 FPS.

 

[InVasMani]

More pcie bandwidth helps when there is destructible terrain and cpu computes the destruction. The updated terrain data can be gigabytes and it takes 20 milliseconds to copy 1 GB on pcie 5.0.

For example in a minecraft clone, you use TNT and blow up half of map. 10 million cubes are destroyed. If 1 cube takes 100 bytes, it updates 1GB. 50 FPS.

That works out to less than half a second on 24GB of VRAM. That's really going to bog you down. I'd hate to think you might lose half a second to completely populate 24GB of VRAM what a major inconvenience.

 

[lexluthermiester]

That works out to less than half a second on 24GB of VRAM. That's really going to bog you down. I'd hate to think you might lose half a second to completely populate 24GB of VRAM what a major inconvenience.

That was such beautiful sarcasm! Top notch!