RX 7900 XTX vs. RTX 4080

[Vya Domus]

It's really not bud.

Do you have anything to back this up, or do you just like refuse to accept this as a fact.

 

[dgianstefani]

Do you have anything to back this up, or do you just like refuse to accept this as a fact.

Unreal Engine 5 Lumen Vs Ray Tracing: Review [2024] |

So, if you're curious about the lumen, what difference it can make, and why we're comparing it to the RTX, keep reading until the end. Unreal Engine 5 Lumen vs Ray Tracing: Which Is Better?

gamersdiscussionhub.com

Read up on the difference between Lumen's hybrid approach and real RT.

Unreal Engine 5 Lumen vs Ray Tracing: Which One is Better?

Epic’s Unreal Engine 4 was one of the most popular game engines of the last generation. It was widely used by AAA giants like EA, Ubisoft, and Microsoft, as well as tons of indie studios. With Unreal Engine 5 and its Lumen and Nanite technologies, Epic is looking to extend its dominance in the...

www.hardwaretimes.com

Lumen vs Ray Tracing: Photorealistic Results in 3D Rendering

Detailed comparison of Lumen & Ray Tracing lighting technologies in Unreal Engine 5 with high resolution interior visualization samples for realistic renders

www.lunas.pro

Lumen is glorified screenspace reflections.

 

[Vya Domus]

Read up on the difference between Lumen's hybrid approach and real RT.

Maybe you should read that. There is no "real" or "fake" RT or whatever else you're implying here.

The only real difference is Lumen has both a software and a hardware accelerated ray tracing code path, which by the way the software implementation isn't even available on PC as far as I know.

Lumen is glorified screenspace reflections.

You're in denial, play the game to see that you're wrong and stop embarrassing yourself with these claims.

 

[dgianstefani]

Maybe you should read that. There is no "real" or "fake" RT or whatever else you're implying here.

The only real difference is Lumen has both a software and a hardware accelerated ray tracing code path, which by the way the software implementation isn't even available on PC as far as I know.



You're in denial, play the game to see that you're wrong and stop embarrassing yourself with these claims.

Ok. Try looking in a mirror in game then tell me it's ray traced bud.

 

[Vya Domus]

Ok. Try looking in a mirror in game then tell me it's ray traced bud.

I have to say there's something so funny about nvidia fanboys believing that unless a game runs like utter crap then it's not using "real" ray tracing.

 

[oxrufiioxo]

People arguing over RT in a thread basically about the OP hoping AMD can do better so he can get an Nvidia card for cheaper is just funny to me.

 

[dgianstefani]

“Lumen” is actually several distinct 3D rendering techniques blended together to form lighting that appears mostly cohesive and, with a decent amount of modern computing power, can produce nearly ray-traced results. Hence why lumen, either software or hardware, is very CPU heavy.

"Hardware" lumen still isn't path traced, it's an approximation that isn't triangle based, using plug in limited versions of real RT, like ray traced shadows, but then, this ideally also requires the integration of nanite based rendering to make up for the shortcomings of using approximation and shortcut, upscaled screenspace "radiance cache", for example.

Unreal Engine 5 goes all-in on dynamic global illumination with Lumen

Lumen is Unreal Engine 5’s new fully dynamic global illumination and reflections system.

www.unrealengine.com

The engine isn't actually "tracing" each frame, it uses a cache over several frames to further downscale "rays" based on blended raster techniques with some RT techniques that can take advantage of dedicated RT hardware. It's certainly impressive considering the performance, but to call Fortnite ray tracing on the same level as something like Cyberpunk is delusional.

There are no magical shortcuts. Lumen "ray tracing" is fast because it's a much simpler model designed to be compatible with older, weaker hardware, such as consoles and AMD GPUs. It's great that it's agnostic and can take advantage of RT cores when they exist, but it's not true RT.

UE6 may introduce actual next gen full RT, by that time consoles should have dedicated hardware, and lumen+nanite is a great compromise way to get better than pure raster/baked lighting for the masses, but it's important to be realistic about what it actually is, instead of taking the marketing at face value and seeing "ray tracing" so it must be the same thing right?

 

[Vya Domus]

There are no magical shortcuts. Lumen "ray tracing" is fast because it's a much simpler model designed to be compatible with older, weaker hardware, such as consoles and AMD GPUs. It's great that it's agnostic and can take advantage of RT cores when they exist, but it's not true RT.

Of course I never claimed it's the craziest most realistic implementation of ray tracing ever but it's disingenuous and straight up false to say that it's somehow not real tracing. The main reason it's so fast is because it's simply using very few rays per pixel. I think it's dumb to disregard it because it can do more with less than other engines who try to do the same.

but to call Fortnite ray tracing on the same level as something like Cyberpunk is delusional.

I said it has the same feature set, which it does, ray traced reflections, AO and GI. I've played both of these games with ray tracing and both of them look bad and far from realistic from time to time, this "it's not on the same level" is just a snobbish nothing argument.

 

[Dr. Dro]

There's no denying that RDNA 2 is deficient at RT workloads. I was having this argument on the 3DMark FSR thread. Speed Way is simply too demanding for this architecture, it takes a fringe halo ultra-high-end design that can chug as much power as it wants, running on third revision, specially binned KXTX type silicon to match an RTX 3070's average score. It may be mighty competent otherwise, but its first-generation RT acceleration capabilities are not very good, I would argue that it's behind even Turing's capabilities in many ways. If it's another hot take of mine, it's again one I'm willing to stand by.

RDNA 3 didn't change that. Hardware Vendor #3 has clearly missed their power and performance targets with the first-gen chiplet design. I'd be happy to let it slide - it's a radical change from the traditional GPU architecture, but knowing full well that it flopped, AMD should rush to be friendlier in pricing, because let's be frank: 7900 XTX is not a remarkable product as far as performance goes, it trade blows with the RTX 4080, has RT performance similar to that of the RTX 3080 Ti or vanilla 3090 despite having 24 GB of VRAM, and it will consume an absurd amount of energy to do so compared to the RTX 4080.

Maybe hardware bugs, maybe drivers broken beyond belief - but something is indeed wrong, because AMD hasn't bothered showcasing anything about the RX 7800 series/Navi 32 silicon - we've got a situation where AMD has no viable product to sell except for the most unpopular segment possible, which is the sub-halo performance segment where GPUs are too expensive for the average buyer, but don't meet the needs of extreme enthusiasts. It's why NVIDIA does things like the Ada product stack, which had the capability to be a genuine generational leap, instead it's sliced and diced to keep the performance per dollar aspect flat or in a very slight decline. Ich liebe kapitalismus... it is what it is.

Regarding compatibility, RDNA 2 is already DirectX 12 Ultimate capable, so it's going to run things, until the driver goes belly up. But the less bleeding-edge techniques that a GPU architecture supports, the worse performance will become. This is an area where they have a significant disadvantage against NVIDIA, especially if they're successful in selling the idea of shader execution reordering and simplified ray intersection to game developers, which you can be confident that they are placing their full weight behind even if it means to treat customers of RTX 20 and 30 series GPUs like second class citizens.

 

[DemonicRyzen666]

There's no denying that RDNA 2 is deficient at RT workloads. I was having this argument on the 3DMark FSR thread. Speed Way is simply too demanding for this architecture, it takes a fringe halo ultra-high-end design that can chug as much power as it wants, running on third revision, specially binned KXTX type silicon to match an RTX 3070's average score. It may be mighty competent otherwise, but its first-generation RT acceleration capabilities are not very good, I would argue that it's behind even Turing's capabilities in many ways. If it's another hot take of mine, it's again one I'm willing to stand by.

RDNA 3 didn't change that. Hardware Vendor #3 has clearly missed their power and performance targets with the first-gen chiplet design. I'd be happy to let it slide - it's a radical change from the traditional GPU architecture, but knowing full well that it flopped, AMD should rush to be friendlier in pricing, because let's be frank: 7900 XTX is not a remarkable product as far as performance goes, it trade blows with the RTX 4080, has RT performance similar to that of the RTX 3080 Ti or vanilla 3090 despite having 24 GB of VRAM, and it will consume an absurd amount of energy to do so compared to the RTX 4080.

Maybe hardware bugs, maybe drivers broken beyond belief - but something is indeed wrong, because AMD hasn't bothered showcasing anything about the RX 7800 series/Navi 32 silicon - we've got a situation where AMD has no viable product to sell except for the most unpopular segment possible, which is the sub-halo performance segment where GPUs are too expensive for the average buyer, but don't meet the needs of extreme enthusiasts. It's why NVIDIA does things like the Ada product stack, which had the capability to be a genuine generational leap, instead it's sliced and diced to keep the performance per dollar aspect flat or in a very slight decline. I love capitalism... it is what it is.

Regarding compatibility, RDNA 2 is already DirectX 12 Ultimate capable, so it's going to run things, until the driver goes belly up. But the less bleeding-edge techniques that a GPU architecture supports, the worse performance will become. This is an area where they have a significant disadvantage against NVIDIA, especially if they're successful in selling the idea of shader execution reordering and simplified ray intersection to game developers, which you can be confident that they are placing their full weight behind even if it means to treat customers of RTX 20 and 30 series GPUs like second class citizens.

RDNA3 sucks at RT, it sucks at RT because it's hardware schedular can't pick which the hell version of it's own "shaders it wants to use" either the "6,144 shaders" or "12,288". It gets into a schedular thrashing after that. Next is the fact they didn't double the amount RT cores. they only added 20% more of compared to the last generation. (80 rt cores vs 96 rt cores.)

 

[Space Lynx]

No, what you're saying is delusional and I'll prove how devastatingly unintelligent your line of thought is since you couldn't come up with an explanation for your nonsensical claims :

View attachment 289919

Look how much faster the 7900 XTX is than a 4090, clearly since this the largest performance delta between these two GPUs in raster performance that I could find then that means TPU's own assessment that the 4090 is 22% faster in raster is irrelevant.

Turns out the 7900 XTX is actually 27% faster than a 4090, obviously, and every other benchmark is wrong/irrelevant/delusional. Would you fully agree with this statement ?

It's truly a shame more games don't use the same engine/way of designing a game as this game does... man those frames are crazy. I could care less about RT, I want to game in OLED 3440x1440 175hz 175 fps ultra with no dips. Fuck RT lol

 

[Vya Domus]

RDNA3 sucks at RT, it sucks at RT because it's hardware schedular can't pick which the hell version of it's own "shaders it wants to use" either the "6,144 shaders" or "12,288". It gets into a schedular thrashing after that.

That's not how any of this works, whether or not the instructions executed by the shaders are dual issue is up to the compiler, Nvidia GPUs have the same problem neither can use all FP32 units at the same time all the time.

 

[JustBenching]

No, you're just wrong.

It's a software implementation only on consoles, on PC is hardware accelerated, it has ray traced reflections, AO and GI, what more do you want ? It's the exact same feature set a game like Cyberpunk 2077 currently has. I know the Nvidia mindshare is devastating but is it really that hard to accept that RT can be done without running like crap on both vendors when it's hardware agnostic ?

Have you actually tried the game in question, fortnite? It uses some form of screenspace reflections, you can tell just by looking at a glass surface. Even your own character doesn't have proper reflections.

 

[Vya Domus]

Have you actually tried the game in question, fortnite? It uses some form of screenspace reflections, you can tell just by looking at a glass surface. Even your own character doesn't have proper reflections.

For the millionth time, you're wrong. Character reflections are indeed broken, probably some weird illumination engine bug but if you pay attention the rest of the geometry in the scene it has proper ray traced reflections.

Look at the 20:05 mark, those are clearly ray traced reflections. Again, I am amused that I have to convince an nvidia fanboy that what they're looking at is indeed ray tracing.

 

[dragontamer5788]

RDNA3 sucks at RT, it sucks at RT because it's hardware schedular can't pick which the hell version of it's own "shaders it wants to use" either the "6,144 shaders" or "12,288". It gets into a schedular thrashing after that. Next is the fact they didn't double the amount RT cores. they only added 20% more of compared to the last generation. (80 rt cores vs 96 rt cores.)

That's really not how any of these GPUs work.

You write a few shaders (usually vertex, pixel, and now Raytracing), those shaders are grouped together into workgroups / blocks / thead groups / whatever you wanna call them. Then in practice, something like 30,000+ of them are shoved into the GPU all at once due to a hyperthreading-like arrangement. IIRC, AMD supports like 40 wavefronts at a time on only 4 physical hardware ALUs.

---------

"Shader count" is becoming weird ever since NVidia decided that one hardware ALU could execute two instructions per clock tick. So NVidia decided that doubled the shader count. Its a bit of twisted logic but whatever, its their marketing department. AMD followed suit as well.

But actual programmers look at either the SM (Symmetric Multiprocessors) on the NVidia chips... or the WGP (Workgroup Processors) on AMD chips. WGP is the smallest unit of the AMD-GPU (starting in RDNA), where one WGP acts kinda-sorta like 2x CUs of old (The Rx Vega was organized into smaller CUs / Compute Units. RDNA and later are organized into WGP that handle twice as much work). IIRC, you can shove something like 40x wavefronts (something like a thread) on a WGP, or maybe 80x ? I'll have to look it up to be sure... but its a lot of wavefronts.

Raytracing still executes inside an SM (on NVidia) or WGP (on AMD). Same with Tensor cores (aka: 4x4 matrix multiplication instructions) on NVidia. The smallest unit is just these SMs or WGPs, which are kinda-sorta like CPU cores but really really weird.

 

[kapone32]

Wow. based on some of the comments I forgot that this is TPU and not Reddit. The narrative is Nvidia and some people will quote Nvidia's narrative in the things that they appreciate. In reality if you put someone in front of a 144hz or higher screen and turn Freesync on they will be hard pressed to tell the difference between the 2 cards an some would be chosen incorrectly. The fact is that Nvidia is better (was) at Crypto Mining and even though that is not relevant now it certainly could be again. Of course people sleep on AMD as the only thing they do is take what Nvidia force feeds into the space and make s a better version of Open source.

Where is the narrative? If you watch Youtube every major Youtber that does a build 9 times out of 10 will use a 4090. When they talk about Multi GPU they talk about SLI without ever mentioning that Crossfire was objectively better when AMD baked it into Adrenilin at the driver level. The worst is people from both Camps but mostly Nvidia that use adjectives like suck, smash, destroy when comparing the cards when the truth is none of that is fair. AMD gets to a 3090 level in Ray tracing and it sucks. Especially (Just like when the 580 was 8 GB) when the current cards give you 20 and 24 GB of VRAM. Yes you need a chip like a 13900K or 7900X3D to push these GPUs. For those that think that 5 GHZ and increased IPC are a joke on AM5.

The 7900X3D forces my 7900XT to use about 2-4 GB more VRAM on every Game vs the 5800X3D. In this world of Ray Tracing and DLSS the best nugget is a 7900XT//XTX will easily drive a 4K 144Hz panel at native and the beauty of Freesync is smooth like butter frame rates, the key to the X3D chips is they boost the frames so that Freesynce is enjoyed more. That is why it feels so smooth. I just finished a 2 stack vs 4 stack 20 unit minimum 100 individuals per unit and it was so smooth I forgot to save the replay. I literally just finished this battle and look at the FPS.

 

[Space Lynx]

I see that the mongrel dogs of the empire are at it again, why must I come in and bring balance to this site, overseeing the mods themself in my divine light!

It's like my last trip to Ohio, ffs. Play nicely citizens of Tamriel!

@kapone32 Don't worry Moon and Star, I am here now

 

[Kovoet]

I've got the dilemma with the ones below. RX 7900XT and the RTX 4070 Ti but everything so far is poiinting to the 4070 Ti for 1440 and two screens

 

[JustBenching]

I've got the dilemma with the ones below. RX 7900XT and the RTX 4070 Ti but everything so far is poiinting to the 4070 Ti for 1440 and two screens

You care about DLSS and RT, you go 4070ti. You don't, you go 7900xt.

Extra points to the 4070ti if you care about power draw, cuda cores, driver stability etc.

 

[AnotherReader]

There's no denying that RDNA 2 is deficient at RT workloads.

RDNA 3 didn't change that.

While you're right about the lackluster ray tracing performance of RDNA 2, RDNA 3 has made significant improvements which are offset to some degree by the lower clocks in ray traced games. Computerbase compared the 7900 XT and 6900 XT by modifying clock speed for matched flops. They found that:

The test series also shows very well that RDNA 3 has increased more in ray tracing than in classic rasterizer graphics. A 9 percent increase in FPS when using the beams doubles to an average of 18 percent in average FPS and 21 percent in percentile FPS.

 

[Vya Domus]

by modifying clock speed for matched flops.

It's not as straight forward as just trying to match the clockspeeds, you'd need to use a GPU debugger to see what's really going on there.

 

[Dr. Dro]

While you're right about the lackluster ray tracing performance of RDNA 2, RDNA 3 has made significant improvements which are offset to some degree by the lower clocks in ray traced games. Computerbase compared the 7900 XT and 6900 XT by modifying clock speed for matched flops. They found that:


View attachment 290331

That's nice and all but it's simply not enough, to the point of being irrelevant. I realize I'm being facetious, but the truth is that this is how the market at large will see it, without idealism or bias supporting the perceived underdog or whatever.

That level of performance barely matches the RTX 3080 Ti and 3090, unfortunately for hardware vendor #3, Ada RT performance is a full generation ahead, and that's being generous. If people can purchase a 4080 for not too much more, and take home a product that has no buts or ifs, you bet that they'll be inclined to do just that.

 

[AnotherReader]

That level of performance barely matches the RTX 3080 Ti and 3090, unfortunately for hardware vendor #3, Ada RT performance is a full generation ahead, and that's being generous. If people can purchase a 4080 for not too much more, and take home a product that has no buts or ifs, you bet that they'll be inclined to do just that.

I wouldn't call a 20% price difference insignificant, but you're right about the lack of RT performance compared to Ada. I'm not sure if that'll change even with RDNA 4 as AMD seems to be sticking to the trend of minimizing the area dedicated to RT.

 

[Dr. Dro]

I wouldn't call a 20% price difference insignificant, but you're right about the lack of RT performance compared to Ada. I'm not sure if that'll change even with RDNA 4 as AMD seems to be sticking to the trend of minimizing the area dedicated to RT.

It's a way of seeing things, but it's $200 for a product that you buy once every 3 years if you're the kind of guy who upgrades every generation, or every 5-6 years if you wait until you can't run ultra settings anymore. And that's assuming optimal, American market conditions - where the 4080 is often seen below MSRP (Micro Center had it at $1150 other day), and the 7900 XTX's $999 price is basically only practiced by AMD themselves (whenever they have stock).

All in all the one thing we can agree on is that the entire market is just rotten. Not to throw shade at anyone in particular, but it's hard to imagine a worse place for the GPU market right now.

 

[AnotherReader]

All in all the one thing we can agree on is that the entire market is just rotten. Not to throw shade at anyone in particular, but it's hard to imagine a worse place for the GPU market right now.

The market is better than it was during the last crypto boom, but that isn't saying much. On a side note, there's now a 7900 XTX available below MSRP: