The Last Of Us Part 2 Performance Benchmark

[kabarsa]

@W1zzard "Ray tracing is not available, which is surprising, considering that the engine looks very similar to the one of Spider-Man 2" what do you mean by that ? Even though Sony have ICE team, which is Initiative for Common Engine, suggesting that TLoU2 and Spider-Man engines are similar is extremely far fetched and probably not true in a lot of ways.

 

[Fluke]

I noted a CPU-load of 80%++ on a Zen3 CPU; that's quite exceptional for a game.
Will there be a CPU follow-up?
I'd be specially interested in a CPU power draw analysis.

Spoiler: There is a CPU benchmark, but there's a catch...

Computerbase did one, but only in their favourite resolution aka 1280x720 Sorry CB, but that's telling me nothing.
In what parallel universe are people using 720p monitors? And why the hell would they couple a 5090 with one?

 

[kabarsa]

Most of Sony’s older game titles prove that RT is simply overhyped. It’s more to give Nvidia an advantage but at the expense of significant performance loss regardless of hardware. I’ve played Horizon Zero Dawn, Forbidden West, God of War and Ghost of Tsushima, and never once did I feel that the lighting and image quality is anywhere subpar without RT.

RT is the right way to universally model many optical effects, you simply describe how much you like offline ray-tracing during lighting bake, and say that real-time ray-tracing is somehow overhyped. Rasterization with a lot of effects use ray-marching for a long time already. Volumetrics with fog and clouds, Cone-Tracing for CryEngine's SVOGI or different kinds of SDF ray marching in UE4 for soft shadows. All this mental gymnastics is essentially many ways to get cheap ray tracing or it's approximation where possible. And Volumetrics and Cone-Tracing were there since PS4 era games, long before nVidia decided to apply hardware acceleration. Why all these techniques try to trace something all the time? Because this is how the actual light model works in geometrical optics. A lot of rasterization techniques with SDF tracing and other stuff are simply dancing around it. There's no way avoiding it if you want to simulate light. RT is not overhyped, this is the right way to do the light. The fact that in the past it was made with offline ray-tracing and now with real-time only shows how good it is. And purely raster stuff like shadow-maps is extremely bad with scaling amount of shadow casters and scene complexity. And scaling is the key. ReSTIR in CP2077 can handle hundreds of shadow casters, can you make a game where each light source can have a dynamic shadow map ??? I can bet you can't, I know it'll scale extremely bad very fast. And before people will say that art style is important and lighting is not, good lighting is working great for stylized Pixal or Dreamworks movies.

 

[W1zzard]

what do you mean by that ?

Same folder structure, same registry keys, same launcher,

Will there be a CPU follow-up?

No plans at this time, too many other things coming up. But I'll probably use it in future CPU reviews (H2 2025)

 

[Deleted member 187435]

I noted a CPU-load of 80%++ on a Zen3 CPU; that's quite exceptional for a game.
Will there be a CPU follow-up?
I'd be specially interested in a CPU power draw analysis.

Spoiler: There is a CPU benchmark, but there's a catch...

Computerbase did one, but only in their favourite resolution aka 1280x720p Sorry CB, but that's telling me nothing.
In what parallel universe are people using 720p monitors? And why the hell would they couple a 5090 with one?

Isn't it normal to do CPU benchmarks at a low resolution to eliminate load on the GPU. The differences there are gigantic. Lower cards like the RTX 4060 may be getting 60fps when paired with a high end CPU but in the real world they'll be paired with much slower or older ones (like my 10th gen i5) so FPS could be halved judging by those results. (10th gen i5 isn't in the chart but the Ryzen 5 3600 is similar level.)

 

[MetroDetroitGaming]

I haven't even been able to get TLOU 1 running for years. Damn game just crashes when starting then just wont open again after. Tried everything I could find online. Sucks

 

[T1beriu]

Computerbase did one, but only in their favourite resolution aka 1280x720p Sorry CB, but that's telling me nothing.
In what parallel universe are people using 720p monitors? And why the hell would they couple a 5090 with one?

This video is for you.

 

[Denver]

Funny how much variation there is between CB and TPU. It seems like their 9070XT is the satanic, hellish-speed edition. Or maybe the scene TPU used is more CPU-bound, and the GPU with lower driver overhead ends up looking better in the graphics; especially since the 9070 and 9070XT perform identically.

 

[Fluke]

I haven't even been able to get TLOU 1 running for years. Damn game just crashes when starting then just wont open again after. Tried everything I could find online. Sucks

Part 1 was a radioactive dumpster fire on release and required dozens of patches.
I recommend verifying ze files, because the game takes 5-10 minutes of shader baking after a patch.

 

[LastDudeALive]

While i appreciate Nvidia for innovating, you gotta see the pattern, hairworks, physx, raytracing, dlss.
They invent some gimmick and push it to game developers until it's a standard, from all the things they did i can only appreciate DLSS, other ones are just to mess with AMD.
We can all agree that if Last of us had ray tracing, the minimum bar for max graphics would be very high and the improvements hard to see.

Hairworks is literally only in 18 games (https://steamdb.info/tech/SDK/NVIDIA_HairWorks/), discounting the modkits, demos, and betas. Only three games of note: Witcher 3, Final Fantasy XV, and Far Cry 4. There is absolutely no way it's a "standard." I've never even seen it discussed outside of Witcher 3.

PhysX is just a physics engine, and runs on the CPU in 99.9% of implementations. Nvidia even made it open-source! There's far more games than you think that use PhysX, it's a solid physics engine, and even the default engine in Unity. The GPU-accelerated effects are entirely OPTIONAL, and would not have been possible at the time (c. 2010) without the work Nvidia put into CUDA. The two choices weren't PhysX effects or something else, the options were PhysX effects or there's literally no other way to run those complex particle effects on the computers of that time. AMD never bothered to create a unified compute language for their GPUs like CUDA, so it's entirely their fault that AMD GPUs couldn't be used for GPU-accelerated effects (or a bazillion professional programs, e.g. Blender). And again, GPU-accelerated effects are optional. If you're referring to dropping 32-bit support for Blackwell, blame the developers for never releasing a 64-bit binary of the game. Nvidia can't be bound to support antiquated code until the end of time just so you can have more smoke in your favorite Arkham game.

Raytracing is how literally all lighting is computed in video games, do you understand that? Baked lighting is just ray tracing done ahead of time by the developer, with lots of tricks and cheats developed over the years to make it look as realistic as possible when the player is in the game and changing the environment. Real-time ray tracing was always going to be the future of video games when GPUs were capable enough, it's entirely AMD's fault that they are asleep at the wheel and still haven't bothered to build a full ray tracing core (lots of the RT pipeline is still done on shaders, even in RDNA 4).

 

[thepath]

I think you should test lower tier GPU on medium or high setting instead of very high. Make no sense to run max setting on budget GPU. I like to see separate test for lower tier GPU running on lower than max setting, so people can know how much fps they will get because most budget GPU users will not be maxing out latest AAA games.

 

[harm9963]

So far GPU usage at 98 to 100 percent , Exclusive full screen and While using Cinematic setting at 4K native TAA ,DLSS isn't needed , indoors way to high fps , with 4K capture , fps outside 80 plus in some spot , 100 to 130 easy , indoors 175 fps , without 4K capture , gain 5 to 10 fps , did OC GPU a little on my next upload , depends on location where GPU is at 100 percent usage , nice bump , but at 100 fps , is more than enough , would not feel the difference , nice for OC benchmarks , but more power and heat .
4090 and 5950X! CPU- RAM 4266 CL18-20-20-20 4X16GB

The Last of Us™ Part II Remastered HDR 4K Native max TAA, NO DLSS

 

[thepath]

5070 Ti is slightly beating 4080 Super despite having less compute power... Either drivers for 5070 Ti is better or maybe this game benefits from extra bandwidth (5070 Ti has higher memory bandwidth than 4080)

 

[bad1080]

how long has it been since we've seen an increase in fps on the relative performance chart?

 

[x4it3n]

CP2077 is number 45 in steam charts as current playing while Monster Hunter Wilds is numbers 6 but no one wants to use it in GPU reviews.

As much as I understand your POV, Cyberpunk 2077 is a lot more stable and reliable than Monster Hunter Wild... CP 2077 was an absolute mess when it released but CD Projekt RED did an amazing job after that. The game can literally run on some potato hardware at the lowest settings. MHW is not bad in any way but it definitely needs some tweaking.

 

[LastDudeALive]

CP2077 was launched in 2020 and is Nvidia's showcase for every new gpu generation, RTX3000, rtx4000 and now RTX5000, people complain about PS5 remasters but CP2077 gets "remastered" every new Nvidia generation, RT overdrive, pathtracing, fake frames...etc.
CP2077 is number 45 in steam charts as current playing while Monster Hunter Wilds is numbers 6 but no one wants to use it in GPU reviews.

MHWilds released barely a month ago and it's receiving constant patches. That's precisely the opposite of what you want in a benchmark suite.

 

[GodMustBeAFK]

No idea why? The 5090 is the most powerful gpu on the market and destroys every other card in pretty much every title. The average is obviously lower, but isn't it like 35% average stronger than the already stupid powerful 4090? 40 in a lot of titles. It's really not that surprising that it's ahead by a substantial amount. Especially with the 5080, which is not much faster than the 4080 at all.

 

[x4it3n]

Last of us part 2 is sitting at #80 on steam player count.
Did anyone fork out 50usd everytime CB2077 got a new update?
Heck The Witcher 3 next gen patch also came out for free...

Part II is a completely different game than Part I so of course they're not going to give it as a free update! And the Part II Remaster is only $10 if you have the original game...

Cyberpunk 2077 is the same game, they just added a DLC for free but the RT Overdrive (Path Tracing) update was not some additional content, just a visual upgrade... TLOU Part I also has a free DLC (included with the game now), and Part II might get one too, so wait and see.

The Witcher III Next-Gen Patch is an update, not a DLC, but the 2 DLCs (Hearts of Stone and Blood and Wine) are not free! So I'm not sure why you're hating on TLOU so much.

5070 Ti is slightly beating 4080 Super despite having less compute power... Either drivers for 5070 Ti is better or maybe this game benefits from extra bandwidth (5070 Ti has higher memory bandwidth than 4080)

TLOU Part II seems very Memory Bandwidth bound yes. Memory Overclocking definitely brings a good performance bump.

MHWilds released barely a month ago and it's receiving constant patches. That's precisely the opposite of what you want in a benchmark suite.

Agree 100%. Another game extremely well optimized and looks amazing too is Split Fiction. I just beat it with my gf yesterday and we had an absolute blast! @W1zzard you should definitely add it to your list! (Even if it's a co-op game)

 

[LastDudeALive]

No idea why? The 5090 is the most powerful gpu on the market and destroys every other card in pretty much every title. The average is obviously lower, but isn't it like 35% average stronger than the already stupid powerful 4090? 40 in a lot of titles. It's really not that surprising that it's ahead by a substantial amount. Especially with the 5080, which is not much faster than the 4080 at all.

Yes, but the performance scaling is more than what's been seen in other games. It's not that it's ahead, it's that it's ahead by that much.

IMO, it's due to the memory bandwidth. You can see cases of otherwise comparable cards (e.g. 4060-Ti vs 3070, or 4070 vs 3080), where the card with more bandwidth gets slightly better performance. And there's a MASSIVE jump in memory bandwidth from the 4090 to the 5090 with a 512-bit bus AND GDDR7.

 

[kabarsa]

Hairworks is literally only in 18 games (https://steamdb.info/tech/SDK/NVIDIA_HairWorks/), discounting the modkits, demos, and betas. Only three games of note: Witcher 3, Final Fantasy XV, and Far Cry 4. There is absolutely no way it's a "standard." I've never even seen it discussed outside of Witcher 3.

PhysX is just a physics engine, and runs on the CPU in 99.9% of implementations. Nvidia even made it open-source! There's far more games than you think that use PhysX, it's a solid physics engine, and even the default engine in Unity. The GPU-accelerated effects are entirely OPTIONAL, and would not have been possible at the time (c. 2010) without the work Nvidia put into CUDA. The two choices weren't PhysX effects or something else, the options were PhysX effects or there's literally no other way to run those complex particle effects on the computers of that time. AMD never bothered to create a unified compute language for their GPUs like CUDA, so it's entirely their fault that AMD GPUs couldn't be used for GPU-accelerated effects (or a bazillion professional programs, e.g. Blender). And again, GPU-accelerated effects are optional. If you're referring to dropping 32-bit support for Blackwell, blame the developers for never releasing a 64-bit binary of the game. Nvidia can't be bound to support antiquated code until the end of time just so you can have more smoke in your favorite Arkham game.

Raytracing is how literally all lighting is computed in video games, do you understand that? Baked lighting is just ray tracing done ahead of time by the developer, with lots of tricks and cheats developed over the years to make it look as realistic as possible when the player is in the game and changing the environment. Real-time ray tracing was always going to be the future of video games when GPUs were capable enough, it's entirely AMD's fault that they are asleep at the wheel and still haven't bothered to build a full ray tracing core (lots of the RT pipeline is still done on shaders, even in RDNA 4).

people often confuse PhysX CPU solver and GPU particle physics or cloth sim. as you said PhysX was the backbone for Unity physics for ages and the same for UE4.
glad to see someone saying that ray tracing was always the end-game, it was always obvious

 

[x4it3n]

Yes, but the performance scaling is more than what's been seen in other games. It's not that it's ahead, it's that it's ahead by that much.

IMO, it's due to the memory bandwidth. You can see cases of otherwise comparable cards (e.g. 4060-Ti vs 3070, or 4070 vs 3080), where the card with more bandwidth gets slightly better performance. And there's a MASSIVE jump in memory bandwidth from the 4090 to the 5090 with a 512-bit bus AND GDDR7.

Yes it's due to Memory Bandwidth. Cyberpunk 2077 w/ PT even get 50% more performance!

The 4090 is hugely memory bottlenecked, let's not forget that it has 60% more CUDA Cores than the 4080 SUPER but is only ~30% faster on Average! There's a lot of performance untapped and it's due to the L2 Cache being crippled to 72MB instead of 96MB on a full AD102 and also because the 4090 should have had 24Gbps chips instead of 21Gbps.
The 5090 only has 96MB L2 Cache (out of 128MB for a full GB202) but it has a 512-bit bus and GDDR7 memory so it allows more Cores to be filled if the L2 Cache gets saturated (but it is power constrained most of the time and Core clocks are pretty much the same as Lovelace too since it's still on TSMC 4nm and not 3nm).
But I would have loved to see a comparison of a 4090 with 96MB and 5090 with 128MB to see the performance impact that it could have had compared to the 4090 et 5090 that we have (the crippled ones).

glad to see someone saying that ray tracing was always the end-game, it was always obvious

All the people who have been in the IT or PC Gaming scene for many years, knew that Ray Tracing and mostly Path Tracing was the Holy Grail of Graphics. But we didn't have the Hardware for it (and still barely have it today... a 5090 having 30fps at Native 4K is not really great). I waa expecting Blackwell to have a much better RT/PT performance due to the supposed "new architecture" but I guess we'll have to wait for Rubin or whatever next architecture they will use for RTX 60s.

 

[Visible Noise]

Btw Monster Hunter Wilds is interesting choice because it's the most unoptimized POS LOL, people still play it regardless

Would you believe that POS has sold over 10 million copies already? I don’t get it (obviously).

 

[LastDudeALive]

glad to see someone saying that ray tracing was always the end-game, it was always obvious

Yes, every person in the entertainment industry as a whole knew ray tracing is the tech of choice. Movies did this two decades ago. Compare Toy Story (baked lighting, reflection maps, etc. i.e. same lighting techniques as games c. 2010) with Cars (which used ray tracing for shadows and reflections). It's miles apart. The problem is just that real-time ray tracing requires orders of magnitude more processing power in orders of magnitude less space. Pixar was using server farms to render 30 seconds of Cars a day, but it's taken 20 years to be able to do the same level of rendering real-time on one GPU.

 

[kabarsa]

Yes, every person in the entertainment industry as a whole knew ray tracing is the tech of choice. Movies did this two decades ago. Compare Toy Story (baked lighting, reflection maps, etc. i.e. same lighting techniques as games c. 2010) with Cars (which used ray tracing for shadows and reflections). It's miles apart. The problem is just that real-time ray tracing requires orders of magnitude more processing power in orders of magnitude less space. Pixar was using server farms to render 30 seconds of Cars a day, but it's taken 20 years to be able to do the same level of rendering real-time on one GPU.

We are in a strange transition stage at the moment, but I am sure that overall progress in ML denoisers and great sampling like in ReSTIR will get us there a lot faster. It's mind blowing what nVidia achieved with ReSTIR in Cyberpunk 2077, and what Lumen can achieve with 0.5 to 2 rays per pixel. Yes, this is temporal accumulation, there're temporal issues sometimes, but still ReSTIR DI + GI in CP2077 runs on 4070Ti+ hardware, and handles complex light sources, emissive surfaces, it all just works. So I think it's closer than we think. The moment we hit that DI + GI plateau, it can scale up to a cinematic levels depending on how much hardware you can throw at it. Initial cost is high, and some of it is fixed cost, but resolution scaling and spp scaling after that is great. I'd argue that CP2077 is already on that plateau, and maybe few others to a lesser degree, like Snowdrop engine in Avatar, or Star Wars Outlaws already uses ReSTIR DI. That transition is so painful because of economic reasons mostly I think, GPU pricing, mining boom, COVID, TSMC bottleneck, PS5 generation slow start, tons of crossgen releases and so on.

 

[LastDudeALive]

We are in a strange transition stage at the moment, but I am sure that overall progress in ML denoisers and great sampling like in ReSTIR will get us there a lot faster. It's mind blowing what nVidia achieved with ReSTIR in Cyberpunk 2077, and what Lumen can achieve with 0.5 to 2 rays per pixel. Yes, this is temporal accumulation, there're temporal issues sometimes, but still ReSTIR DI + GI in CP2077 runs on 4070Ti+ hardware, and handles complex light sources, emissive surfaces, it all just works. So I think it's closer than we think. The moment we hit that DI + GI plateau, it can scale up to a cinematic levels depending on how much hardware you can throw at it. Initial cost is high, and some of it is fixed cost, but resolution scaling and spp scaling after that is great. I'd argue that CP2077 is already on that plateau, and maybe few others to a lesser degree, like Snowdrop engine in Avatar, or Star Wars Outlaws already uses ReSTIR DI. That transition is so painful because of economic reasons mostly I think, GPU pricing, mining boom, COVID, TSMC bottleneck, PS5 generation slow start, tons of crossgen releases and so on.

To be honest, I hope Nvidia (and AMD, but I don't trust them to be smart enough to do this) radically re-engineers their allocation of space with their next generation. I would honestly be totally fine if Nvidia stopped adding more CUDA cores in the next generation, and just devoted 10x more space to RT cores. That's really what's needed. Rasterization has hit the limit, it's obvious. Nvidia's architecture is still fundamentally stuck at Turing. CUDA cores take up 90% of the space, with RT and Tensor cores "bolted on." We need far more RT cores.