PCI-Express 4.0 Performance Scaling with Radeon RX 5700 XT

[SASBehrooz]

Thanks for review. i was going to build PCIe 4.0 system and choosing Radeon RX 5700 XT Over the RTX 2070 Super. now im sure i go with RTX 2070 Super
thanks

 

[tfdsaf]

Thanks for review. i was going to build PCIe 4.0 system and choosing Radeon RX 5700 XT Over the RTX 2070 Super. now im sure i go with RTX 2070 Super
thanks

The RX 5700XT is still a much better value just in terms of raw GPU performance, its only 6% slower than the 2070 super, but up to 30% cheaper. hardwareunboxed did a 39 games test and found the RX 5700xt just 6% slower on average.

Plus PCI-E 4 is more for AMD's CPU's and the infinity fabrik, as well as much faster SSD's, plus you are future proofing. But yeah, even a RTX 2080ti won't outbandwidth a pci-e 3 slot.

 

[computerdiehard]

The only true test is to put that video card on an older chipset. It's not a true test all on the same motherboard.

 

[W1zzard]

The only true test is to put that video card on an older chipset. It's not a true test all on the same motherboard.

Why?

 

[computerdiehard]

Why because the X570 has 44 PCI-Xpress lanes and it not like you can go into the BIOS and set how many PCI-Xpress lanes you want to dedicate. You also can't set what PCI version you want in the BIOS.

Correct?

 

[W1zzard]

Why because the X570 has 44 PCI-Xpress lanes and it not like you can go into the BIOS and set how many PCI-Xpress lanes you want to dedicate. You also can't set what PCI version you want in the BIOS.

Correct?

Actually no. The PCIe speed (version) can be set in the BIOS, the number of lanes can be changed by physically taping off some on the slot, the card will autonegotiate to the correct link width.

 

[Jo3yization]

Please test bandwidth at 4K PCIe 3.0 vs 4.0 with ReLive recording, I've been able to reproduce consistent spaced stutter when recording with ReLive under high VRAM usage in games like SottR & Gears 5, but dont have a PCIe 4.0 platform to compare against. I'm running a Sandisk Ultra II SSD & tried multiple drivers to see if that may have been the issue but it makes no difference, if I bring the Vram usage down by lowering texture quality or simply turn Relive off the stutter disappears so it's definitely connected to VRAM & ReLive.

 

[Athlonite]

Please test bandwidth at 4K PCIe 3.0 vs 4.0 with ReLive recording, I've been able to reproduce consistent spaced stutter when recording with ReLive under high VRAM usage in games like SottR & Gears 5, but dont have a PCIe 4.0 platform to compare against. I'm running a Sandisk Ultra II SSD & tried multiple drivers to see if that may have been the issue but it makes no difference, if I bring the Vram usage down by lowering texture quality or simply turn Relive off the stutter disappears so it's definitely connected to VRAM & ReLive.

What CPU are you using

 

[Jo3yization]

What CPU are you using

i7-6700k,
Z170A Gaming Pro
2x8gb ddr4 3200mhz cl16 (Double checked in CPU-Z, its in dual channel, passes memtest no problems, I dont get any crashes or BSODs).
Sapphire pulse RX 5700(non-XT).

Here's what the stutter looks like;;

For the entire first ~26mins of the video recording is smooth, but after a few cutscenes at the timestamp, watch the vram usage + frametimes; you can see as it climbs to about 7.3gb when the stutter starts I knew what was happening straight away so I lowered texture streaming to eliminate it, it's happened in a bunch of other titles over different driver revisions too(19.9.1&2,, 19.11.2>3 and 19.12.1>2), . If ReLive is off, it doesnt happen at all even at just under 8gb usage. I also noticed it tends to be more prevalent in DX12 API, if I use DX11 in borderlands 3 for example at 3440x1440 same settings etc, the vram usage was a good 2gb lower so the problem didnt occur.

Here's Borderlands 3(this was on 19.11 drivers)

First benchmark run is in DX11 API, there was a bug with 19.11 that reports incorrect VRAM usage in some utilities so the numbers arent reliable but they do show higher usage in DX12 at the very least, but you can see me run the benchmark smoothly in DX11 at first, then swap to DX12 & immediate stutter, which I resolve by lowering settings to lower the VRAM usage.

So as you can see I can consistently reproduce the problem & I've sent the details to AMD, though recording at 3440x1440 is a bit of a niche using the RX 5700 so obviously not many people have encountered it yet. Seeing as relive *should* be GPU encoding with only a small load on CPU, for it to happen only when VRAM usage is high I'm guessing theres either some kind of memory leak or texture swapping going on around a certain level of usage(looks like about 7.3gb) in Gears 5, as ReLive itself must need a certain amount for encoding & its somehow conflicting with the game.

So far I've reproduced it in:
Borderlands 3
Jedi Fallen order
Escape from tarkov
Red dead redemption 2
need for speed heat
& gears 5.

I did theorize maybe its a disk issue, so I tried changing the recording location to my other SSD along with dropping bitrate from 50MB/s to 5MB/s, and swapping from HEVC that I normally use to simply AVC but none of it made any difference, recording is smooth until the game VRAM usage gets to a certain point.. I also tried running completely stock(as you can tell Im running a core OC on the RX 5700) but it also did not resolve the problem. Also plenty of free space on all 3 of my SSDs & they all record smoothly at the same bitrates before the stutter occurs or if I simply turn texture quality//other settings down to keep the vram usage from hitting whatever amount triggers it.


For now I've just learned to live with it and put it down to either driver or ReLive/GPU limitations when high resolution recording ..But thought..Maybe PCIe 4.0 can help after seeing the behaviour of the 5500 XT driven into its bandwidth limit at igorslab; https://www.igorslab.de/en/amd-rade...h-the-morepowertool-and-into-bandwidth-limit/ even though its due to a much smaller bus width causing the issue with that GPU, maybe under certain high VRAM usage scenarios it could occur with Navi 10 GPUs under PCIe 3.0 as well...

Or its all happening exclusively on the GPU with the Relive encoder fighting game textures for VRAM which makes the most sense & might be fixable via driver updates to prioritize or 'balance' VRAM usage better by dynamically reducing bitrate at high levels of usage to prevent texture/encoder swapping for lack of a better term.

 

[Athlonite]

Or its all happening exclusively on the GPU with the Relive encoder fighting game textures for VRAM which makes the most sense

I'd say you just hit the proverbial nail on the head right there. So by turning down the textures from Ultra to High you don't get the stutter well that seem to point directly at what you were getting at it's a resource fight between ReLive and the game engine for VRam personally for just recording I'd turn down the textures to high but leave it on Ultra for pure game play time

 

[gphantom]

No one who is benchmarking video cards seem to use XPLANE or other flight sim software in their tests. Xplane 11 especially, doesn't rely on cpu cores but rather how fast a cpu and graphics card is. It does not rely on linked cards either so, to show a graphics card's true speed and bandwidth capability, use XPLANE 11 as a testing platform. THEN we can truely judge how capable a graphics card and even a motherboard/cpu/memory really is.

Phil
Application/Software engineer since 1976

 

[EarthDog]

No one who is benchmarking video cards seem to use XPLANE or other flight sim software in their tests. Xplane 11 especially, doesn't rely on cpu cores but rather how fast a cpu and graphics card is. It does not rely on linked cards either so, to show a graphics card's true speed and bandwidth capability, use XPLANE 11 as a testing platform. THEN we can truely judge how capable a graphics card and even a motherboard/cpu/memory really is.

Phil
Application/Software engineer since 1976

what?

The point here is to see if games are effected by available slot bandwidth. While a single core game may rely on clock speed, what makes you think there is more data going through the pcie slot that makes 'all flight sims' better for this type of test? How do you think it will differ from the other, mostly negligible, results?

 

[Jo3yization]

So.. This thread popped up in my emails and I figured I'd post an update.

The stutter is definitely VRAM usage and can occur even *without* Relive running in certain titles that are very resource hungry, one such title is CoD: Warzone that many users are reporting stutter & fps drop problems with, I narrowed the issue down to the amount of VRAM the game uses as opposed to it's 'in-game' calculator showing less than half the estimate.

Here's a vid pretty much demonstrating the cause of stutter many are experiencing with CoD warzone, its basically texture swapping stutter, but the big factor is CoD warzone uses enough VRAM to cause swapping at only 1440p/1080p on lower VRAM cards, it'll hit 4gb usage very easily, I even tested very low 720p and was still able to hit 4gb vram usage.

Just sharing since its interesting as even though the cause of high VRAM usage is game related, the symptom is basically the same as the ReLive//VRAM usage stutter. Which confirms it definitely is some texture swapping issue.

 

[ProDigit]

in other words, what you should have done as additional testing is pcie 4.0 x16, pcie 4.0 x8, pcie 4.0 x4, and pcie 4.0 x1. And see the performance differences there.

 

[Jo3yization]

Oh I would if I had the hardware to test, but I'm stuck on a 6th gen PCIe 3.0 system for the moment. I also encountered the same High VRAM//swapping issue on Horizon Zero Dawn and was able to eliminate 'periodic' stutter on my 8gb RX5700 using the same method that works with warzone, lowering VRAM usage, it seems once the game has to start swapping data between the Video Memory and system ram, stutter occurs.. But given the bandwidth of both VRAM & DDR4 @ 3200mhz(40GB/s), the limitation in this case would actually be PCIe 3.0 bandwidth(32GB/s max),, so in theory, PCIe 4.0 with double the bandwidth would scale higher with faster memory speeds & possibly eliminate stutter at some point. I wish I had a PCIe gen 4 setup to test.

This VRAM 'buffer' performance loss when VRAM limits are reached also lines up with the Horizon Zero Dawn 'PCIe gen 3.0 vs 4.0' test done by Techspot shown in the benchmarks listed here: https://www.techspot.com/review/2084-amd-or-intel-for-gaming-benchmarking/ , you can actually see that at the higher usage of 1440p & 4K, the PCIe gen 4 system is performing better in HZD with an 8GB GPU than on the otherwise faster intel system limited to Gen 3.0 that dominates pretty much all other benchmarks.

Here's a link to my own High VRAM usage stutter test in Horizon Zero Dawn that anyone running a Navi GPU can likely replicate too;

 

[msroadkill612]

The point here is that we sheep have, (since pcie2 gpus anyway) for many years been overspending on gpuS & wasting 8x pcie lanes.

even on pcie 3, 16 lanes rarely helps.

the upshot is u can have a; cheaper to build, as good, 8x pcie 4 GPU, yet still be as future proof as a 16 lane pcie 3 gpu for when games inevitably do start finding more uses for evolving/faster/bigger system resources.

 

[Jo3yization]

We can? I think it really depends on use case. Some situations & games will definitely benefit. Especially newer titles even at 1080p.

 

[B_Bang]

Any chance you could do this on the 6600 NON - xt card. Given it's lower performance i think it might be more affected by PCIe version...

 

[ExalyThor]

Thanks for the testing! I still use an i7 2600K and I wondered if the pcie 2.0 will be a bottleneck.