Horizon Forbidden West Performance Benchmark

[b1k3rdude]

Is this test system reasonable? It seems madly overclocked to me. 330W and 6 GHz P cores? This is a chiller territory. I cannot imagine how that 280 AIO can cool this.

Yeah I thought @W1zzard must be running that CPU at intel recomended settings, as the AF280 I beleive only supports upto 250w?

"a bit better than the modern RDNA 3 models (for example RX 6800 XT beating RX 7800 XT"

"The game is also DRM-free and works fine completely offline."

- Not suprising given the PS5 is based on AMD silicon, but its still nice to see the venerable 6800XT making its younger sibling look stupid. The 6000 series are still better bang for buck when ist comes to raster perf in a lot of gasmes, even now.

- this is very good news, and is basically the gold standard of what ALL bloody games should be.

 

[Denver]

The GCD on Navi31 and 32 are onm TSMC N5. It is the APUs that are on N4.
As for the performance, I think it is more likely that AMD over estimated their software teams' ability to utlitise the dual-issue FP32 on RDNA3.
Also the reduction in Infinity Cache size does not help. Navi21 has 128MB while Navi31 only has 96MB. Navi32 is down to 64MB that is less than the 6700 XT that had 96.
Navi31 has a 384-bit memory bus and significantly faster memory to compensate, Navi32 basically got nothing vs Navi21.

In fact, the implementation of dual-issue FP32 cannot be done via a driver, it has to be implemented on the developers' side. If consoles had this capability, it would probably be more common for games to run well on RDNA3 hardware without frequent driver intervention.

 

[BoggledBeagle]

Yeah I thought @W1zzard must be running that CPU at intel recomended settings, as the AF280 I beleive only supports upto 250w?

Arctic liquid freezer II 240 I had could cool 300W without the CPU hitting max temperatures, I just tested 360 version and with 253W load the CPU temp was 73°C.

I believe that 360 or 280 which is comparable to 360 can cool 330W comfortably.

But if he really overclocked all P cores to 6 GHz, that may require such voltage that the CPU may be hitting quite high temperature.

I think it could work with partial gaming load, but it just surprised me when I saw it.

 

[Johnthedon1]

RT reflections would have been nice, and added a lot to the visual presentation.

Not a lot to reflect tho in the jungle, global illumination would have been a better tough

 

[starfals]

Woa, 7900XTX sure is close to 4090... and sure is kicking the 4080's backside. In both price and performance. I mean, 4090 here is literally double in price. DOUBLE for that extra 10 fps? No thanks Nvidia. Also, that card is not only passing the 4080, but its also a lot cheaper than it here. That new so called 1000 bucks price is a myth in my country. Prices start at 1200 to 1450! YES 1450!! European prices in my country have gone absolutely insane. Stores scalping, Nvidia BS, Euro tax... thats how a 1k card becomes 1300 plus. Recently the 7900 XTX got a few price cuts too, so yeah. I dunno what to say. I don't wanna buy AMD, I'm too used to Nvidia but yeah. Anyone that doesn't care about the things i care about (Ansel, GF experience, recording and where settings are) should seriously ignore Nvidia for high end gaming imho, at least the people that want to save some cash. Is raytracing even that important? I never turn it on, and we had that for like 3 generations of Nvidia video cards. I got a 3080 and a 2080. It just tanks the performance too hard, and changes the art style of some games too. Its not always better to turn it on. I did a blind test with a few friends, they actually preferred the non RT version. It had less shadows and darkness. I did like it in Minecraft, but again... it changes the point of the game. It was made to be a simple game lol, not better than Crysis (and hoboy, RTX sure makes it gorgeous) Sometimes i do mind how RT looks, its not just darkness either. Even in Spiderman, i dislike the reflection in the windows. The old ugly one actually looks better to me while flying around the city. I sure ain't paying double the price of a 7900XTX for something that i don't always use/care about, and the sweet sweet 10-25 extra fps. I mean if the 4090 offers a bit more performance, why does it cost double in my place lol. Why not a little bit more? Ye ye, its the best, so they made it cost whatever they wanted.

P.s. Woa, this game runs like a dream, and looks amazing. A rare thing in 2024.
P.s2. What i wrote came from an Nvidia fanboy btw, i never had an AMD GPU, but holy cow. They sure look good on paper
P.s.3. Yep, DLSS is amazing, i will probably miss it if i ever went to AMD, so obviously im just waiting for RTX 5080. Again, if people don't care about DLSS/fake frames ( FSR 3 works on 3080, f off Nvidia) or all my little things, AMD sure looks better, on paper, as i said. If we are all rich, sure? Let's all get a 4090, that is the best video card after all. Start by buying me one first tho ;p

Oof, seems I will need to upgrade my GPU, just, not this generation.

Luckily I have many other titles to catch up on. I enjoyed Death Stranding, kept me busy for weeks. Next up, Control.

Ha, same here actually! I just finish DS, and im about to try out Control, after Forbidden west of course This generation just sucks, if i will pay these crazy prices, might aswell wait for 5080 and pay for that. If 4080 is already 1200-1400 now, even with the so called price reduction, what are they gonna do? make 5080 around 1600-1700? Fine, at least that will be worth the money. Gives me more time to save up too!

 

[Luminescent]

Found a bug!
Game has a memory leak or something that degrades performance after a while, everything works great with framerates over 60fps and i enter a town and dips to 40-50 fps.
Restart the game and framerate jumps back to 60-70 fps, no change in weather/lightning or anything, this might only happen on AMD so i have no knowledge if it happens on Nvidia.
At this point all benchmarks based on latest driver from AMD and this game version are not final, if this is fixed numbers might change.

 

[b1k3rdude]

and sure is kicking the 4080's backside. In both price and performance.

Yeah but not on power useage, its why in the end I went for a 4080.

Found a bug!

Why you should always wait for atleast 1-2 patches before buying a game.

 

[Vya Domus]

In fact, the implementation of dual-issue FP32 cannot be done via a driver, it has to be implemented on the developers' side.

This definitely works on the driver side, when shaders get compiled they can include dual issue instructions, no one has to hand optimize shaders for this to work. That may be necessary but it's not mandatory.

Dual issue ALUs just don't make that much of a difference, it was the same when Ampere was released and it had double the FP32 performance yet games didn't ran twice as fast and still don't, nor ever will.

 

[Godrilla]

These are GPU tests so the CPU needs to be as small a bottleneck as possible to isolate GPU performance.

14900K is a native 6GHz part and while it can peak at 330W in all-core loads, it usually consumes less than 200W while gaming because no game comes close to saturating it. No probs for a 280mm AIO.

Am I calculating this right at 19cents per Kwh Google result for NY rates for my locallity. The delta rate per hour is 2.85 Cents ( at 150 watts extra usage) before adding in the air conditioning rate hidden cost.
If you use the PC at 4 hours daily X 7 days, X 52 weeks X 3 years the total will be $124 before taxes and fees.

This is worse case scenario and the averages seem about 2/3 lower so the real out of pocket cost before taxes and fees ( as well as air conditioning) is probably around ( $124 X 2/3) is $83 after 3 years 4 hours per day 7 days a week usage. Yay numbers.

 

[Luminescent]

Seems the memory leak is happening on Nvidia too, found a long reddit post with people having bad framerates in dialogues or if they stay in menu changing settings, at this point game is fu...ed, we wait a month or two to be fixed.
Why launch a game in this state.

 

[Denver]

This definitely works on the driver side, when shaders get compiled they can include dual issue instructions, no one has to hand optimize shaders for this to work. That may be necessary but it's not mandatory.

Dual issue ALUs just don't make that much of a difference, it was the same when Ampere was released and it had double the FP32 performance yet games didn't ran twice as fast and still don't, nor ever will.

But the idea isn't to double performance. From what I've read, dual-issue capability only works for certain operations, and is more complex to implement than on Nvidia GPUs, so it requires interference from capable hands more often.

"Unfortunately, testing through OpenCL is difficult because we’re relying on the compiler to find dual issue opportunities. We only see convincing dual issue behavior with FP32 adds, where the compiler emitted v_dual_add_f32 instructions. The mixed INT32 and FP32 addition test saw some benefit because the FP32 adds were dual issued, but could not generate VOPD instructions for INT32 due to a lack of VOPD instructions for INT32 operations. Fused multiply add, which is used to calculate a GPU’s headline TFLOPs number, saw very few dual issue instructions emitted."

"It relies heavily on the compiler to find VOPD possibilities, and compilers are frustratingly stupid at seeing very simple optimizations. For example, the FMA test above uses one variable for two of the inputs, which should make it possible for the compiler to meet dual issue constraints. But obviously, the compiler didn’t make it happen."

"On the other hand, VOPD does leave potential for improvement. AMD can optimize games by replacing known shaders with hand-optimized assembly instead of relying on compiler code generation. Humans will be much better at seeing dual issue opportunities than a compiler can ever hope to. Wave64 mode is another opportunity.,"

So yes, a well-tuned implementation would make a difference.

Microbenchmarking AMD’s RDNA 3 Graphics Architecture

Editor’s Note (6/14/2023): We have a new article that reevaluates the cache latency of Navi 31, so please refer to that article for some new latency data.

chipsandcheese.com

 

[Prince Valiant]

Seems the memory leak is happening on Nvidia too, found a long reddit post with people having bad framerates in dialogues or if they stay in menu changing settings, at this point game is fu...ed, we wait a month or two to be fixed.
Why launch a game in this state.

There's no incentive not to as long as people keep pre-ordering or buying close to launch.

 

[Luminescent]

I have to restart the game every 20-30 minutes, this is stupid, refund on the way if it can be done.

 

[sixor]

I miss cpu corea charts (how many cores does the game uses) it was cool info, useless today but interesting

 

[Onasi]

I actually prefer the settings on medium/high. Very high just looks "meshy forced" not sure how else to describe it. lol that is happening a lot with modern games I noticed, but I can always turn settings down a notch, it will look better plus I gain fps, so meh it is what it is.

Glad to see I am not alone on this. A lot of modern titles striving for realism look overdone to me, especially on their highest settings. I think I started noticing it in the previous gen titles like Fallout 4 which had this uncanny “over-tessellated” look and these days, yeah, “over-meshed” does seem like a good way of describing it. I look at some of the screenshots in the article and it just looks overwhelming to me. Maybe it’s personal preference (it definitely is), but nowadays I strongly prefer a clean, defined art-style in games to the photo-realism race which feels like throwing more and more on the screen. That first UE5 tech-demo game, Immortals of Aveum, just looked like a mess of effects and particles, for example.

 

[Vya Domus]

.

Those benchmarks prove my point, all GPUs have comparable FMA op/s despite RDNA2 not being dual issue. Ada and RDNA3 are dual issue yet they have similar FP32 add performance for example, once again showing Nvidia has no real advantage in their implementation of dual issue ALUs.

 

[Hxx]

Pretty cool that the 7900xtx is significantly cheaper and performs better than a 4080. Big win for AMD gamers

 

[QuietBob]

I miss cpu corea charts (how many cores does the game uses) it was cool info, useless today but interesting

I wouldn't call this kind of information useless

 

[napata]

Those benchmarks prove my point, all GPUs have comparable FMA op/s despite RDNA2 not being dual issue. Ada and RDNA3 are dual issue yet they have similar FP32 add performance for example, once again showing Nvidia has no real advantage in their implementation of dual issue ALUs.

Nvidia does not have dual issue ALUs. They run into the same problem in regards to scaling but it's misinformation to call it dual issue. A 4090 does actually have 16384 FP32 ALUs and not 8192 with dual issue.

On mid/lower end cards this does actully show an advantage for Nvidia (just look at the 4080 vs 7900XTX) but the 4090 just sucks at utilizing its shading power and has terrible scaling even compared to a 4080.

 

[Solaris17]

Game is fun, the intro and flash back dialogue to zero dawn is a bit cheesy, but the environment is gorgeous and the controls are easy to pick back up. It doesnt appear they changed them at all which is great. Performance has been fine, but you can almost tell its a port. Horizon object and texture popnin is a thing if you are a console player you know what im talking about and that translates to the PC version. Though even from elevated places the world below you is remarkably detailed, which is nice.

 

[Vya Domus]

Nvidia does not have dual issue ALUs. They run into the same problem in regards to scaling but it's misinformation to call it dual issue. A 4090 does actually have 16384 FP32 ALUs and not 8192 with dual issue.

On mid/lower end cards this does actully show an advantage for Nvidia (just look at the 4080 vs 7900XTX) but the 4090 just sucks at utilizing its shading power and has terrible scaling even compared to a 4080.

Whether something is dual issue or not you still need the same number of functional FP32 units, so Navi31 has 12288 FP32 units, it doesn't make sense to say something "actually has X number of FP32 units" because they're gonna have the same number of ALUs regardless.

It's hard to know exactly what type of instructions Nvidia cards issue because they don't disclose their ISA and the compiler is closed source but look at the benchmark again, most FP32 operations have similar throughput, this suggests both types of compute units issue instructions in a similar manner.

 

[Denver]

Those benchmarks prove my point, all GPUs have comparable FMA op/s despite RDNA2 not being dual issue. Ada and RDNA3 are dual issue yet they have similar FP32 add performance for example, once again showing Nvidia has no real advantage in their implementation of dual issue ALUs.

"compilers are frustratingly stupid at seeing very simple optimizations. For example, the FMA test above uses one variable for two of the inputs, which should make it possible for the compiler to meet dual issue constraints. But obviously, the compiler didn’t make it happen."

In fact, the analysis reveals that despite suboptimal optimization by the compiler, certain operations still leverage the FP32 dual-issue capability in RDNA3. Further manual optimization would uncover additional opportunities to exploit dual-issue capabilities.

 

[W1zzard]

I wouldn't call this kind of information useless

View attachment 340482View attachment 340483

GameGPU doesn't actually test these, the numbers are approximated using math

 

[QuietBob]

GameGPU doesn't actually test these, the numbers are approximated using math

Oh well
I didn't know that. Thanks a lot for the info!

 

[heraq]

Why not test the RX 7600 XT?