News Reviewer reports RTX 5080 FE instability — PCIe 5.0 signal integrity likely the culprit

[beyondlogic]

Excellent launch Nvidia! Watch AMD still make no real attempt to capitalize on.

dont know how there suppose to capitalize on it you want the amd horse to stumble out the gate as well lol.

 

[bit_user]

Excellent launch Nvidia! Watch AMD still make no real attempt to capitalize on.

They can't, because drivers/FRS4 aren't ready. The only thing worse than failing to capitalize on this situation is if they did launch now and had embarrassing problems of their own.

 

[coromonadalix]

that only means, WAIT 1 year before buying things now, hoping all bug will be out and addressed ... this is getting so ..... still happy with my rtx 3060 loll

seeing this from 4000 series connectors fiasco, 5000 series is no better, ill continue to wait

 

[snemarch]

And the general consensus of people that have taken the time to test it is that PCIe 4.0 x16 offers the same performance as PCIe 5.0 x16 on the 5090. Maybe when I have time (next month, or maybe March?) I'll try to do my full test suite in PCIe 4.0 and 3.0 mode. But right now I have a lot of other GPUs to test!

Probably not a whole lot of reason to do that – for one, other people have already done it, and... what do people expect?

Once the bulk of textures have been loaded, how much data do you need to transfer on a frame-per-frame basis?

If you can find a game that constantly streams in huge textures, or some computational workloads that can't fit the entire dataset in GPU memory, sure – otherwise, your time is probably better spent on more interesting topics

 

[Pierce2623]

They can't, because drivers/FRS4 aren't ready. The only thing worse than failing to capitalize on this situation is if they did launch now and had embarrassing problems of their own.

My point is they show no real interest in putting a genuine effort in on discrete GPUs. They develop graphics IPs for APU big or small. Discrete cards are just further monetizing of their APU tech.

 

[Pierce2623]

dont know how there suppose to capitalize on it you want the amd horse to stumble out the gate as well lol.

My point is they’re not even trying to capitalize.

 

[Pierce2623]

Haha! Agreed.
That's along the lines of, "any machine can be a smoke machine if operated poorly enough!"

All electronics actually run on smoke rather than electrons. They continue to work until you let the smoke out.

 

[adamXpeter]

So, PCIe 5.0 ended up being worse than a pointless waste of money - it's downright harmful!

You have missed this part, even higher performance nVidia card ran in that config:

For reference, this same setup was used to benchmark GPUs like the RX 7900 XTX, RTX 4080, RTX 4090, and even the RTX 5090, with no issues whatsoever.

 

[beyondlogic]

My point is they’re not even trying to capitalize.

With what product mate they can't capitalise on it if it's not ready sending a duff product out early hurts market share. They would if they could have by now if it was ready. Just be patient and wait this is exactly why games etc come out half broken

 

[bit_user]

Probably not a whole lot of reason to do that – for one, other people have already done it,

The latest I'm aware of is TechPowerUp testing a RTX 4090 with a i9-13900K. The faster your CPU and GPU, the more of a bottleneck PCIe should become. @JarredWaltonGPU , I'd like to see PCIe scaling tested using a 9800X3D and RTX 5090 at PCIe x16 3.0 vs. 4.0 vs. 5.0.

I don't see a lot of value in testing many cards, the way TechPowerUp did. Maybe throw in a couple other RTX 5000 cards. 5070 should be interesting, due to being a 12 GB card.

Once the bulk of textures have been loaded, how much data do you need to transfer on a frame-per-frame basis?

I'm sure games with larger levels are continually streaming in assets. That's the whole point of DirectStorage, in fact.

​

You could cheat, a little bit, and see which games got a greater benefit from PCIe 4.0 in their testing. Focus more (but not exclusively) on those:

NVIDIA GeForce RTX 4090 PCI-Express Scaling with Core i9-13900K

We take a second look at PCI-Express performance scaling of NVIDIA's GeForce RTX 4090 flagship card. This time with a Core i9-13900K, which hopefully helps us highlight more differences than what we've seen with Ryzen 5800X last year. We've also added minimum FPS and ray tracing testing.

www.techpowerup.com

otherwise, your time is probably better spent on more interesting topics

I think PCIe scaling is interesting!

I also thought the RTX 4090 power scaling Jarred did was interesting.

https://www.tomshardware.com/news/improving-nvidia-rtx-4090-efficiency-through-power-limiting
​

Heck, I think memory scaling is plenty interesting.

https://www.techpowerup.com/review/intel-core-ultra-9-285k/27.html
​

 

[bit_user]

My point is they show no real interest in putting a genuine effort in on discrete GPUs. They develop graphics IPs for APU big or small. Discrete cards are just further monetizing of their APU tech.

I don't know how you can say that, when the RX 6950X managed to beat the RTX 3090 Ti!

Source: https://www.tomshardware.com/reviews/amd-radeon-rx-6950-xt-review/6
​

Just because RDNA3 was more lackluster in its results doesn't mean AMD hasn't been trying. Let me assure you that there was nothing lazy about their GCD/MCD chiplet approach! That was a lot bolder than anything Nvidia has done, recently. And it's not something that carries over to their APUs.

My point is they’re not even trying to capitalize.

They launched Ryzen 9000 too early, seemingly trying to capitalize on Raptor Lake's woes, and look where that got them!

It's more important for them to launch fully-baked products, than to pounce on a moment of weakness in their competitor, because you only get one chance to make a first impression. Those impressions and launch reviews will stick around for a products entire sales cycle and people will continually refer back to them. So, the most important thing is to have a solid launch.

 

[JTWrenn]

I don’t know that I’d blame the retailers with all the rumors of EXTREMELY limited stock. Large retail outlets were talking about receiving single digits of 5090s. They probably got 2 5080s.

I think the answer is a bit of both.

https://videocardz.com/newz/scalper...force-rtx-50-founders-edition-sales-in-europe

Looks like the scalpers were doing their evil bs and there was very low quanitites.

 

[bit_user]

You have missed this part, even higher performance nVidia card ran in that config:

I saw where they ran the RTX 5090 @ PCIe 5.0. Just because some cards can work at that speed doesn't invalidate my point.

 

[Peksha]

I'd like to see PCIe scaling tested using a 9800X3D and RTX 5090 at PCIe x16 3.0 vs. 4.0 vs. 5.0.

I already wrote about this: the bus can severely limit performance at x8 (for z600-800 Intel and x870 with gen 5 nvme) width and simultaneous use of DLSS/FSR/XESS "performance"/"ultra performance" scaling, especially for the 1080/1440 Ultra performance scenario.
Also, no one has yet shown whether the RTX50 uses the full bw pcie 5.0 - for this, you need to do a gpgpu aida test for reading and writing to memory. Because it seems that it cannot, since Nvidia has historically had problems with bandwidth of PCIE (image from c&c, rdna3 review)

 

[snemarch]

I'm sure games with larger levels are continually streaming in assets. That's the whole point of DirectStorage, in fact.

Yeah, which is why I added the "If you can find a game that constantly streams in huge textures" part – a large open-world game that does seamless loading would be interesting (Ratchet and clank? Cyberpunk? Old stuff like Warcrack?)

But it would need titles specifically selected for that, and a method that comes up with a meaningful benchmark, imho FPS probably isn't the right metric for it.

What I wanted to avoid is another review that benchmarks games that do classic asset pre-loading, then concludes there's hardly any gain for the (relatively!) small frame-by-frame datasets.

 

[adamXpeter]

I saw where they ran the RTX 5090 @ PCIe 5.0. Just because some cards can work at that speed doesn't invalidate my point.

So, one card not running is validate your point?

 

[spongiemaster]

The latest I'm aware of is TechPowerUp testing a RTX 4090 with a i9-13900K. The faster your CPU and GPU, the more of a bottleneck PCIe should become. @JarredWaltonGPU , I'd like to see PCIe scaling tested using a 9800X3D and RTX 5090 at PCIe x16 3.0 vs. 4.0 vs. 5.0.

I don't see a lot of value in testing many cards, the way TechPowerUp did. Maybe throw in a couple other RTX 5000 cards. 5070 should be interesting, due to being a 12 GB card.

Already been done. Not that hard to check their website.

 

[bit_user]

I already wrote about this: the bus can severely limit performance at x8 (for z600-800 Intel and x870 with gen 5 nvme) width and simultaneous use of DLSS/FSR/XESS "performance"/"ultra performance" scaling, especially for the 1080/1440 Ultra performance scenario.

Where? Why would that be?

Also, no one has yet shown whether the RTX50 uses the full bw pcie 5.0 - for this, you need to do a gpgpu aida test for reading and writing to memory. Because it seems that it cannot, since Nvidia has historically had problems with bandwidth of PCIE

What transaction size does that test use, or is it just timing a single transfer of that size? Were multiple concurrent transactions tried? I could certainly believe their driver or PCIe controller isn't pipelining, since game engines might generally overlap enough transfers that it's not necessary for them to.

 

[spongiemaster]

I already wrote about this: the bus can severely limit performance at x8 (for z600-800 Intel and x870 with gen 5 nvme) width and simultaneous use of DLSS/FSR/XESS "performance"/"ultra performance" scaling, especially for the 1080/1440 Ultra performance scenario.

It's debatable whether even PCIe 1.1 is "severely limiting performance."

 

[bit_user]

But it would need titles specifically selected for that, and a method that comes up with a meaningful benchmark, imho FPS probably isn't the right metric for it.

IMO, fps is all you can really measure. Asset streaming performance could vary the amount of draw-in or popping textures, but that's not something easily measurable outside of the engine.

What I wanted to avoid is another review that benchmarks games that do classic asset pre-loading, then concludes there's hardly any gain for the (relatively!) small frame-by-frame datasets.

I get that, but the point of reviews is to guide readers about the importance of PCIe speed on their gaming experience. If most games simply do pre-loading, then it's fair to include some of those in the test suite so you're not left with an impression that PCIe speed is more important than it really is.

 

[bit_user]

So, one card not running is validate your point?

If it's like 0.1% of all RTX 5000 cards, then yeah we just call it bad luck and move on. Especially if they accept it for exchange, under warranty.

If it turns out to be a significant proportion of the RTX 5000 cards, then I'd say it's a problem.

 

[Peksha]

Where? Why would that be?

The lower the resolution, the more transactions are made cpu-gpu - it just fills the channel. This is very important for eGPU - so to minimize losses on excess transactions for playing with it, it is advisable to enable the highest resolution possible. Otherwise, the losses will be too high. I had such a setup in the past.
So in a situation where you have only 8 lanes left after connecting all your ssds, and you are actually rendering in 360/720/1080p, this can have an extremely negative impact on performance, up to negative efficiency

 

[bit_user]

Already been done. Not that hard to check their website.

If you're going to embed the image, then a link to the article would've been courteous.

Also, you just picked one resolution and average perf; not 1% minimums, which you didn't even acknowledge. So, people here are going to think that's the final word on the matter. Not to mention other questions, like the test setup, which games & settings were used, etc.

Here's the 1% minimum for 1080p:

Source: https://www.techpowerup.com/review/nvidia-geforce-rtx-5090-pci-express-scaling/32.html
​

Going from 16 GB/s to 32 GB/s (e.g. PCIe 3.0 x16 -> PCIe 4.0 x16) yields an improvement of 2.6%.
Going from 32 GB/s to 64 GB/s (e.g. PCIe 4.0 x16 -> PCIe 5.0 x16) yields an improvement of 1.4%.

Their ray tracing benchmarks showed no real benefit, at any resolution:

https://www.techpowerup.com/review/nvidia-geforce-rtx-5090-pci-express-scaling/30.html
​

The gains with DLSS 4 were only about 1 fps per step, for 16 GB/s -> 32 GB/s -> 64 GB/s:

https://www.techpowerup.com/review/nvidia-geforce-rtx-5090-pci-express-scaling/31.html
​

Considering the data @Peksha provided on the RTX 3070 vs RX 7900 XTX, it seems worth also testing that AMD card (just up to PCIe 4.0 x16, obviously).

For the curious, also doing some synthetic tests to understand whether Nvidia's PCIe controller has bottlenecks preventing it from utilizing the full PCIe bandwidth would be ideal. Doesn't 3D Mark have some test like that?

 

[bit_user]

It's debatable whether even PCIe 1.1 is "severely limiting performance."

Again, you're taking their data out of context. This data is weird, since it doesn't align with their raster data on CP2077. They gave this explanation:

"For this test I selected Cyberpunk 2077. The test scene is not our standard test scene from the reviews, but a slightly different location, which is a bit more practical to test in this scenario, so numbers aren't directly comparable. We're also testing with RT set to max."

Source: https://www.techpowerup.com/review/nvidia-geforce-rtx-5090-pci-express-scaling/31.html
​

The last sentence seems to be the key, because it does line up pretty well with their 3840x2160 framerates on CP 2077 with RT:

https://www.techpowerup.com/review/nvidia-geforce-rtx-5090-pci-express-scaling/30.html
​

For the sake of comparison, here's how raster performance scaling (sans DLSS) looks.

https://www.techpowerup.com/review/nvidia-geforce-rtx-5090-pci-express-scaling/8.html
​

At higher resolutions, it's less responsive to PCIe speed than their aggregates. This makes me wonder if maybe CP 2077 wasn't the best choice as a test vehicle for DLSS 4. Especially because enabling DLSS should have a similar effect as dropping to 2560x1440 (native).

See, there's quite a lot of nuance to all of this. You're reaching for simple answers, but performance is often too complex for that to give the full picture. There's a reason they ran several hundred tests and wrote a 34 page article. You're not going to just summarize the whole thing in one graph. People who care should go and read the piece for themselves.

 

[snemarch]

IMO, fps is all you can really measure. Asset streaming performance could vary the amount of draw-in or popping textures, but that's not something easily measurable outside of the engine.

Yeah, I can't think of relevant measuring points, or how to present it to users in a meaningful way – which actually makes this somewhat interesting 😉. If there's no mix of System / DirectX APIs that that make sense to generically monitor, I do wonder if it would be worth finding a few select game titles and instrumenting them specifically.

For a classical preload game you could get a precise "time for assets to load and enter level" (rather than using a stopwatch), which might be of some value... for a streaming game it's harder to come up with quantifiable metrics.

I get that, but the point of reviews is to guide readers about the importance of PCIe speed on their gaming experience. If most games simply do pre-loading, then it's fair to include some of those in the test suite so you're not left with an impression that PCIe speed is more important than it really is.

Most definitely – if interesting outliers are found and benchmarked, the results should either note that "for your average titles the performance difference is within measurement error", or include a number of those normal titles. I'm annoyed by the constant push towards BIGGER NUMBER BETTER and the marketing departments + influencers convincing people it'll make a meaningful difference to them...