Hogwarts Legacy Benchmark Test & Performance Analysis

[AusWolf]

Intel RT scores have been updated, I benched RT Low by accident. I'm still including RT Low for RX 7900 XTX, RTX 4090 and Arc A770 as they are an interesting data point. RTX 4090 RT low results testing right now, will be added within the hour

Thanks for all the work you're doing on this! TPU is really the only tech site where something like this happens.

It's interesting to note that the 7900 XTX is 3.25x faster when RT is on Low vs Ultra, while the A770 is only 1.44x faster (at 1080p). So with RT on Ultra, the 7900 XTX is 0.88x faster (that is, slower), while with RT Low, it's 1.98x faster than the A770. I wonder why this is. There must be something in the Ultra setting that's missing, or different at Low, and kills AMD GPUs.

 

[Xuper]

wait what ? RT now fixed on 7900XT ? it was around 6.2 fps at 4K , now 18.3fps?

 

[AusWolf]

wait what ? RT now fixed on 7900XT ? it was around 6.2 fps at 4K , now 18.3fps?

Huh? Where?

 

[Xuper]

Just look at 7900 XTX 6 fps , above 6900 XT , also i'm talking about minimum

 

[AusWolf]

Just look at 7900 XTX 6 fps , above 6900 XT , also i'm talking about minimum

You must be looking at something else.

 

[W1zzard]

RTX 4090 RT Low has been added

It's interesting to note that the 7900 XTX is 3.25x faster when RT is on Low vs Ultra, while the A770 is only 1.44x faster (at 1080p). So with RT on Ultra, the 7900 XTX is 0.88x faster (that is, slower), while with RT Low, it's 1.98x faster than the A770. I wonder why this is. There must be something in the Ultra setting that's missing, or different at Low, and kills AMD GPUs.

Correct. This is clear evidence that something is wrong on AMD RT Ultra

 

[AusWolf]

RTX 4090 RT Low has been added

So, the 4090 is 1.58x faster with RT Low at 4K compared to RT Ultra. If we assume that the same is true of the 4080, then it would achieve 40.5 FPS at 4K with RT Low, which is 5.9 FPS, or 17% faster than the 7900 XTX.

Now, where did I see the same number before? @fevgatos, @Vya Domus?

 

[JustBenching]

So, the 4090 is 1.58x faster with RT Low at 4K compared to RT Ultra. If we assume that the same is true of the 4080, then it would achieve 40.5 FPS at 4K with RT Low, which is 5.9 FPS, or 17% faster than the 7900 XTX.

Now, where did I see the same number before? @fevgatos, @Vya Domus?

Not in cyberpunk or control or portal, for sure

 

[Vya Domus]

Now, where did I see the same number before?

It must have been your imagination, everyone knows the 4080 is 50% faster.

 

[AusWolf]

Not in cyberpunk or control or portal, for sure

Definitely. But the point is, you can't say that the 4080 isn't 17% faster than the 7900 XTX in any scenario.

My new theory is that there must be some limitation in AMD's RT engine that only allows something like a certain maximum number of rays, or a maximum something to be calculated at the desired speed, and it cracks when the game needs more. So for example, when you have x or 2x or 3x, etc. number of rays, the 4080 is consistently 17% faster, but when you exceed a certain number, AMD cards just drop dead.

 

[Xuper]

ok so wizzard changed result. he added low rt to chart and this made me confused.

Wizzard, I think you should add Tab for image gallery like this :

Creating Tabs with HTML, CSS & jQuery – Inspirational Pixels

In this tutorial you'll learn how to easily code tabs using HTML, CSS and jQuery for use in your web projects.

inspirationalpixels.com

 

[JustBenching]

Definitely. But the point is, you can't say that the 4080 isn't 17% faster than the 7900 XTX in any scenario.

My new theory is that there must be some limitation in AMD's RT engine that only allows something like a certain maximum number of rays, or a maximum something to be calculated at the desired speed, and it cracks when the game needs more. So for example, when you have x or 2x or 3x, etc. number of rays, the 4080 is consistently 17% faster, but when you exceed a certain number, AMD cards just drop dead.

Well, let's both thank computerbasede cause it made my point much easier with their testing. Check this out

No RT, the cards are equal. RT reflections ON , the difference is 8%. WOW, only 8%, amd really got close to nvidia in RT performance didn't they? Then he tests RT Ambient occlusion. Difference now is 17% (the one you are claiming). Then he activates shadows, now the difference is a whooping 40%!! Then he activates all of them, now the difference is 46%!!!!

Why does the difference keep growing the more RT effects he adds if the difference between the 2 cards is "17%"?

 

[Karath]

Whats up with 7900 XT VS 7900 XTX with RT on? Something fishy going on.

*edit*
Oop, RT low on the XTX in the same chart. Nevermind, though it definitely looks out of place.

 

[AusWolf]

Well, let's both thank computerbasede cause it made my point much easier with their testing. Check this out

View attachment 283214

No RT, the cards are equal. RT reflections ON , the difference is 8%. WOW, only 8%, amd really got close to nvidia in RT performance didn't they? Then he tests RT Ambient occlusion. Difference now is 17% (the one you are claiming). Then he activates shadows, now the difference is a whooping 40%!! Then he activates all of them, now the difference is 46%!!!!

Why does the difference keep growing the more RT effects he adds if the difference between the 2 cards is "17%"?

Hmm... that kind of confirms my theory, with shadows being the culprit of tanking AMD cards. Everything's sort of fine until RT shadows are turned on.

I guess AMD will have to figure out what's with RT shadows that eats their GPUs alive for RDNA 4.

It reminds me of the old VLIW times when AMD was crap at tessellation, but OK at everything else. Now, RT shadows seem to be their kryptonite.

 

[gffermari]

When we compare performance in RT, we test in pure RT games/benchmarks or games with RT reflections, shadows, AO or GI.

We don’t care if that’s not the case in real games as we don’t care if no one plays games at 720p with a 13900/7950 when we test cpu performance.

We care about the performance difference and that’s why we create “lab” conditions to test.

Radeons 7000 RT performance is acceptable in general but bad for the asking price.

On topic, the fps numbers don’t look nice, but using DLSS-FSR they’ll be great for most of the cards.

 

[Vecix6]

As I can see game is now ready for all gamers so... is there some kind of 0-day patch?

 

[Spinnek]

N3M3515 said:
Something i would like to know also, is there any tool to know exactly how much vram is being used?

Nothing that's accessible to us as far as I know

Just one moment, guys

Have you seen this?
https://www.reddit.com/r/nvidia/comments/j1tm2t
In short, you should be able to verify, how much VRAM is being used in Afterburner (4.6.3 Beta 2 Build 15840 or later):
Near the top and next to "Active Hardware Monitoring Graphs" click the "..."

Click the Checkmark next to "GPU.dll", and hit OK

Scroll down the list until you see "GPU Dedicated Memory Usage", "GPU Shared Memory Usage", "GPU Dedicated Memory Usage \ Process", "GPU Shared Memory Usage \ Process"

Pick and choose what you want to be tracked using the checkmarks next to them. "GPU Dedicated Memory Usage \ Process" is the # that most closely reflects the # we find in FS2020 Developer Overlay and Special K (DXGI_Budget, except Unwinder uses D3DKMT api)

Click show in On-Screen Display, and customize as desired.

***

I'm using this "GPU Dedicated Memory Usage \ Process" monitoring for a week, it shows lower usage than previous VRAM allocation

 

[londiste]

Port Royale is the closest thing to a "pure aryan tracing benchmark" and the difference between the 4080 and 7900XTX is no where near 50%. So as usual you are speaking utter nonsense. Only a hardcore fanboy would assume this one game is a better representation of RT performance rather than a literal benchmark specifically designed to showcase differences in RT performance.

Not quite. Port Royal is a hybrid RT benchmark, similar to how games were expected to be using DXR. Specifically, Port Royal does RT reflections and RT shadows. While the effects are used heavily throughout the scene this might be simpler to GPUs than heavy RT games. Particularly, games that do RT lighting or AO - think Control, Cyberpunk or now Hogwarts.

And for pure ray tracing benchmark, for games you would need to look at either Quake 2 RTX or Portal with RTX.

 

[Avro Arrow]

Howcome the game uses 14GB vram at 1440p but the trx 3080 is not crippled by it?

I asked the same thing and the reason is that the chart says "VRAM Usage" when it should say "VRAM Allocation". I'm glad that I'm not the only one who was confused by that!

And for pure ray tracing benchmark, for games you would need to look at either Quake 2 RTX or Portal with RTX.

Yeah, jeez, Portal RTX... Every time I saw that game getting looked at, it just laid waste to everything that wasn't an RTX 4090!

 

[Vya Domus]

Specifically, Port Royal does RT reflections and RT shadows. While the effects are used heavily throughout the scene this might be simpler to GPUs than heavy RT games. Particularly, games that do RT lighting or AO - think Control, Cyberpunk or now Hogwarts

It's the other way around, AO and GI are typically the least computationally expensive RT effects because you do not need a lot of rays to get an approximate result, while reflections and shadows require a lot more rays to get a convincing result.

Of course developers can be stupid about it (or do it on purpose) and increase the number of rays where there is no visual pay off, RTAO is particularly redundant, most of the time it's indistinguishable from SSAO but if you want to you can still tank the performance with it.

 

[Dakula74]

Game use VRAM 12gb dedicate + 12gb alocated. 1440p

 

[AusWolf]

When we compare performance in RT, we test in pure RT games/benchmarks or games with RT reflections, shadows, AO or GI.

We don’t care if that’s not the case in real games as we don’t care if no one plays games at 720p with a 13900/7950 when we test cpu performance.

We care about the performance difference and that’s why we create “lab” conditions to test.

Radeons 7000 RT performance is acceptable in general but bad for the asking price.

On topic, the fps numbers don’t look nice, but using DLSS-FSR they’ll be great for most of the cards.

I'm not questioning the validity of "lab condition" testing. I'm questioning the usefulness of it.

I asked the same thing and the reason is that the chart says "VRAM Usage" when it should say "VRAM Allocation". I'm glad that I'm not the only one who was confused by that!

It says "VRAM usage" so that less tech-savvy readers can also have a glimpse at what's going on. Besides, you cannot really test how much VRAM you're actually "using".

 

[QuietBob]

Just one moment, guys

Have you seen this?
https://www.reddit.com/r/nvidia/comments/j1tm2t
In short, you should be able to verify, how much VRAM is being used in Afterburner (4.6.3 Beta 2 Build 15840 or later):
Near the top and next to "Active Hardware Monitoring Graphs" click the "..."

Click the Checkmark next to "GPU.dll", and hit OK

Scroll down the list until you see "GPU Dedicated Memory Usage", "GPU Shared Memory Usage", "GPU Dedicated Memory Usage \ Process", "GPU Shared Memory Usage \ Process"

Pick and choose what you want to be tracked using the checkmarks next to them. "GPU Dedicated Memory Usage \ Process" is the # that most closely reflects the # we find in FS2020 Developer Overlay and Special K (DXGI_Budget, except Unwinder uses D3DKMT api)

Click show in On-Screen Display, and customize as desired.

***

I'm using this "GPU Dedicated Memory Usage \ Process" monitoring for a week, it shows lower usage than previous VRAM allocation

Thanks for sharing and the detailed step-by-step instructions!

I've noticed that the same statistics are also tracked by "Memory usage" and "Memory usage / process" without the plugin. The second number represents the VRAM utilized by the running game, whereas the first one takes into account everything else running in the background, i.e. other apps, browser tabs, Windows UI, etc. "Memory usage" represents the total amount of VRAM utilized at the moment and will always be greater.

Total VRAM usage can also be monitored in HWinfo with the "GPU D3D memory dedicated" sensor.

 

[JustBenching]

I'm not questioning the validity of "lab condition" testing. I'm questioning the usefulness of it.

If you don't run "lab conditions" then you end up thinking that a pentium 4 is as fast as a 13900k cause they get the same framerate in 8k.

The usefulness of lab conditions testing is, the moment a game shows up (like cyberpunk or hogwarts or control), that shows a 50% difference in RT performance, we don't blame drivers or optimizations cause the difference is supposed to be "17%". We know the difference isn''t supposed to be 17% and the 50% we are seeing is perfectly normal

 

[londiste]

RTAO is particularly redundant, most of the time it's indistinguishable from SSAO but if you want to you can still tank the performance with it.

I wholeheartedly disagree. RTAO is very distinguishable and so far always results in a clearly more realistic look for the game than other AO methods.