Forza Horizon 3 Benchmark Thread

[bearxor]

Looks like SLI is dead to the world.

Yeah, doesn't look like my dual R9 390's are going to really do me any good here...

 

[4K_shmoorK]

Just had 15GB of 49GB downloaded and the windows store presented me with an error and started my download over. Thanks Obama

 

[IllogicalGlory]

Digital Foundry's verdict is that the RX 480 is tied with the GTX 1060 on ultra settings without MSAA aside from potential asset streaming issues ...

Forza Horizon 3 is arguably a big step backwards compared to Forza Motorsport 6: Apex in terms of performance on AMD GPUs. Polaris microarchitecture got very lucky when AMD decided to increase geometry throughput at a whim ...

I'd say Forza Horizon 3 presents a good use case where DX12 doesn't help AMD much and that's in cases with extremely geometry limited scenarios where Nvidia has an advantage regardless ...

OoO raster (this needs to be exposed in DX12 soon), GPU compute cluster culling, async compute, triangle ordering, and these other sorts of optimizations are a god send for AMD GPUs when they are known to struggle a lot with geometry processing ...

If geometry is the problem, why is the 970 almost as bad as the 390?

 

[SPBHM]

If geometry is the problem, why is the 970 almost as bad as the 390?

almost as bad? it looked around 40% faster than the 390.

but yes, compared to the new cards they are all slow.

the 390 specially looks like a disaster, I don't see how it should perform this much worse than a 480.
hopefully it's like Doom, a patch greatly boosted the 390 performance, but, once again the old products are left behind and waiting for drivers/patches, another reason to avoid old used cards

 

[IllogicalGlory]

You're right, it was much worse, but the 970 is still nowhere near the 480 or 1060, and it is better at geometry than 480.

My guess is that the devs had a 480 and a 1060 for testing and they only did QA for those archs and that we might see patches or driver updates fixing the 390 and 970.

 

[railven]

Recent DF video makes me wonder if the DX12 commentary some have made may have some truth to it. Do devs/AMD/NV have to go back and patch DX12 titles's specifically for older gen hardware? If this is true, this doesn't bode confidence in me that cards would be good for as long as they were during DX11.

I'm sure AMD would fix it for R9 390. Might also explain some of the Fury numbers I've seen.

 

[Bacon1]

http://arstechnica.com/gaming/2016/...ormance-framerate-and-4k-tests-in-windows-10/

Going to say it appears to be CPU related, as people are having better luck @ 4k 30fps locked vs 1080p 60fps.

Since the game was designed for 30fps makes sense that the engine might have some issues going to 60+.

Considering some of the funky results reviewers have found and some reviews having cards do great while another has it doing terrible, gotta be something other than the GPUs alone. CPU, Memory, ?? as the bottleneck.

 

[ThatBuzzkiller]

If geometry is the problem, why is the 970 almost as bad as the 390?

Geometry processing IS an issue with AMD! Why else does Fiji get slaughtered by an RX 480 or a GTX 1060 ?! (Both of which have higher geometry performance.)

RX 480 and the Fury X feature the same ISA too so there's no difference to be had in the internal driver shader compiler either!

You're right, it was much worse, but the 970 is still nowhere near the 480 or 1060, and it is better at geometry than 480.

My guess is that the devs had a 480 and a 1060 for testing and they only did QA for those archs and that we might see patches or driver updates fixing the 390 and 970.

The RX 480 is the one with higher geometry performance than the 970 but the gap is just further enhanced with much higher compute power ...

AMD needs to learn that there's more to games than just pure compute performance and if they raise core counts on Vega without scaling geometry throughput then AMD is in for another tough time like they were for Fiji ...

It's great that AMD has lot's of influences in D3D12 and it's especially true if we take a look at the shader model 6 preview but AMD needs to start decisively winning in the current games badly with margins of around 5-10 % in the product range it's competing in ...

 

[Bacon1]

Only gamegpu shows fury doing bad, the other site shows it just under 980 ti.

they have the 480 underclocked which is why its so close to the 470 and under 1060.

Will be nice for more sites to look into it and do some in depth testing, as something strange is going on

 

[IllogicalGlory]

I'm not saying geometry isn't a problem on AMD, I am questioning whether it is the issue in this game.

The graph you linked shows it only winning by 10% and that benchmark is obviously affected by other factors as well.

http://www.hardware.fr/articles/951-8/performances-theoriques-geometrie.html

In these synthetics, for example, it is getting ripped by Pascal and Maxwell and it's barely above the 390, et. al.

The biggest gains come when there is a lot of geometry and MSAA is turned on. In DF's tests, MSAA is off.

 

[Azix]

This is something I have been wondering about. When people talk about geometry performance, they usually point to tessellation benchmarks. but geometry is not tied to that is it? Tessellation has its own separate bottleneck apart from polygons on screen, no?

IMO claiming a geomtry issue outside of tessellation in this game wouldn't make sense. Unless the trees and road are very high polygon. Is it known that the game uses heavy tessellation?

The game seems to have other issues besides that.

http://www.hardware.fr/articles/951-8/performances-theoriques-geometrie.html

that link from above seems to address more actual polygon performance, but how many polygons do actual games push realistically to hit a bottleneck?

Geometry processing IS an issue with AMD! Why else does Fiji get slaughtered by an RX 480 or a GTX 1060 ?! (Both of which have higher geometry performance.)

RX 480 and the Fury X feature the same ISA too so there's no difference to be had in the internal driver shader compiler either!



The RX 480 is the one with higher geometry performance than the 970 but the gap is just further enhanced with much higher compute power ...

AMD needs to learn that there's more to games than just pure compute performance and if they raise core counts on Vega without scaling geometry throughput then AMD is in for another tough time like they were for Fiji ...

It's great that AMD has lot's of influences in D3D12 and it's especially true if we take a look at the shader model 6 preview but AMD needs to start decisively winning in the current games badly with margins of around 5-10 % in the product range it's competing in ...

 

[Sweepr]

Another website showing poor performance from pre-Polaris AMD GPUs in Forza Horizon 3. Fury X below RX 480 level, GTX 1060 faster than both:

www.computerbase.de/2016-09/forza-horizon-3-benchmark/2/#diagramm-forza-horizon-frames-langsamer-als-60-fps-in-3840-2160

 

[ThatBuzzkiller]

I'm not saying geometry isn't a problem on AMD, I am questioning whether it is the issue in this game.

The graph you linked shows it only winning by 10% and that benchmark is obviously affected by other factors as well.

http://www.hardware.fr/articles/951-8/performances-theoriques-geometrie.html

In these synthetics, for example, it is getting ripped by Pascal and Maxwell and it's barely above the 390, et. al.

The biggest gains come when there is a lot of geometry and MSAA is turned on. In DF's tests, MSAA is off.

It pretty much is an issue in this game. Aside from the immediate mode tiled rasterization approach which can save Nvidia GPUs tons of overdraw, that's the only thing I can think of where Nvidia has a serious advantage is in geometry performance unless Forza Horizon 3 was using NVAPI like the entire UE4 engine ...

This is something I have been wondering about. When people talk about geometry performance, they usually point to tessellation benchmarks. but geometry is not tied to that is it? Tessellation has its own separate bottleneck apart from polygons on screen, no?

IMO claiming a geomtry issue outside of tessellation in this game wouldn't make sense. Unless the trees and road are very high polygon. Is it known that the game uses heavy tessellation?

The game seems to have other issues besides that.

http://www.hardware.fr/articles/951-8/performances-theoriques-geometrie.html

that link from above seems to address more actual polygon performance, but how many polygons do actual games push realistically to hit a bottleneck?

Tessellation benchmarks do have some issues when extracting geometry performance since it does depend on how fast the tessellation units can expand vertex data but by now I do not believe it to be an issue anymore on AMD since they improved upon it via vertex reuse and primitive discard accelerator on the level where they can match Maxwell microarchitecture's throughput. Testing geometry performance in vertex data limited scenario's isn't ideal either since testing could be limited by video memory or memory bandwidth ...

Tons of games can hit geometry performance bottlenecks on AMD, believe it or not otherwise AMD wouldn't be advertising async compute gains so much. Filling a G-buffer, shadow map rendering, and depth only pre-passes are ALL highly dependent on geometry performance! These are rendering passes where Nvidia is KNOWN for excelling!

 

[Azix]

Tons of games can hit geometry performance bottlenecks on AMD, believe it or not otherwise AMD wouldn't be advertising async compute gains so much. Filling a G-buffer, shadow map rendering, and depth only pre-passes are ALL highly dependent on geometry performance! These are rendering passes where Nvidia is KNOWN for excelling!

That does not follow. We should just wait and see what happens with the game in the future. Seems even people on 1070 GPUs are having issues.

 

[dogen1]

That does not follow. We should just wait and see what happens with the game in the future. Seems even people on 1070 GPUs are having issues.

Yes it does. One of the main reason AMD gains to much from Async. Compute is because their much larger geometry bottleneck gives them a larger window of opportunity to gain from running non geometry bound shaders.

 

[Azix]

Yes it does. One of the main reason AMD gains to much from Async. Compute is because their much larger geometry bottleneck gives them a larger window of opportunity to gain from running non geometry bound shaders.

so microsoft is allowing this and kronos is supporting this simply to cover up specifically AMDs issues?

 

[ThatBuzzkiller]

so microsoft is allowing this and kronos is supporting this simply to cover up specifically AMDs issues?

Yes! More specifically it was Microsoft who called the shots ...

 

[RussianSensation]

Another website showing poor performance from pre-Polaris AMD GPUs in Forza Horizon 3. Fury X below RX 480 level, GTX 1060 faster than both:

Per the author of GameGPU, either the game and/or driver issue exists on AMD GPUs under VHQ. Performance is flying under HQ settings.

Comparing R9 290/290X to GTX780/780Ti, the GPU generation those cards were meant to compete against, Kepler is not even on the map. BTW, at VHQ 1080p, none of these cards are playable which means HQ 1080p appears to be reasonably playable for the R9 290. What do we get?

$399 R9 290 = 53 fps min / 63 fps average
$699 GTX780Ti = 39 fps min / 43 fps average

...Let's conveniently ignore that a $700 card from the same generation loses to a $400 AMD card at HQ settings where the R9 290 is performing well. Now the R9 290 user can sell that card for $100, add $300 saved from not wasting it on the 780Ti turd, and buy him/herself a GTX1070

Interesting to note that Computerbase's findings were that even GTX1080 could not manage perfectly locked 60 fps at 1080p with MSAA in the game...just saying that not all performance issues here are strictly hardware related.

R9 380X is also destroying GTX960 4GB, while RX 470 is enroaching on the GTX980, and beats GTX1060. Without a doubt the RX 480 4GB would have beaten the GTX1060 without much effort. It's par for the course that no one in this thread decided update the thread with this data because I suppose it went counter to all the AMD-bashing. GameGPU even noted in the first review that VHQ settings reduced performance on AMD cards by 2-2.5X but no one bothered reading/translating the review.

@ 1080p, moving from VHQ to HQ:

Fury X moves from 42 fps to 103 fps (!)
RX 470 moves from 43 fps to 75 fps (!)
R9 290X moves from 24 fps to 69 fps (!)

vs.

Tessellation benchmarks do have some issues when extracting geometry performance

That's a nice theory for this title, except for one detail -- GTX770 has higher geometry performance than R9 285/R9 380/R9 380X, R9 290/290X. Yet, R9 380X 4GB beats GTX770 4GB in every single setting in this title, VHQ or HQ, both at 1080p and 1440p.

Also, how do you explain GTX960 consistently beating GTX770 if you argue that geometry performance is limiting GPUs in this title? At 1080p HQ, the 960 is a whopping 37% faster than GTX770 4GB; and still 26% faster at 1080p VHQ (29 fps vs. 23 fps).

Not to mention that GTX780Ti has similar tessellation performance to the GTX980, but even at HQ 1080p, R9 380X is beating 780Ti by 33%.

 

[Shivansps]

There something that is VERY, VERY VRAM intensive at VHQ for sure.

Where are the 4GB VRAM defenders now??????? fist Deux ex, now this... 4GB is getting out of the high quality game petty fast.

 

[Shivansps]

I hope AMD will bring a new driver soon for Hawaii/Fiji and not make an NVIDIA stunt with older GPUs.

Some day you ill have to realise that is just a myth, and Nvidia problems on old cards are mostly VRAM limited, not saying its the only issue, but lack of VRAM is playing a mayor role, same thing gona happen to old AMD cards if VRAM usage is forced to go up again, if 4GB is rendered useless for Ultra quality, you can forget about Furys, its gona fell faster than a 2GB GTX770.

 

[Vaporizer]

There something that is VERY, VERY VRAM intensive at VHQ for sure.

Where are the 4GB VRAM defenders now??????? fist Deux ex, now this... 4GB is getting out of the high quality game petty fast.

4GB cards may fall off the cliffs, but somehow new gen Pascal 3GB cards are perfectly fine for most people. Could be Trump logic...

 

[Bacon1]

There something that is VERY, VERY VRAM intensive at VHQ for sure.

Where are the 4GB VRAM defenders now??????? fist Deux ex, now this... 4GB is getting out of the high quality game petty fast.

Its not 4GB:

470 4GB handles the game fine @ 1080p Ultra MSAA 4x with 37.79 min, 49.7 avg and 75.33 max.

Now the 3GB 1060 bombs hard.

https://youtu.be/l0XejmBCuEk

The 470 4gb even handles 4k @ avg of 31 fps, which means you could do locked 30fps well on it.

Now he said that he had higher framerates when telling the game to run @ locked 60 instead of unlimited, so maybe there is an engine bug there.

Also when Digital Foundry tested they found even Titan XP couldn't handle 1080p 4xMSAA Ultra @ 60fps minimum and still had dips and stutters. Yet they could do 4k w/o MSAA Ultra 60 min w/o issue.

DF also noticed that the 8GB 390 was having massive issues while the 970 3.5/.5 wasn't as bad (but still under either Polaris/Pascal), so again doesn't seem to be VRAM related except for the 3GB 1060.

 

[moonbogg]

Game downloading now. Hope it runs OK. I'll ditch AA and tweak settings. Game looks fun.

 

[Sweepr]

Per the author of GameGPU, either the game and/or driver issue exists on AMD GPUs under VHQ. Performance is flying under HQ settings.

ComputerBase tested at 1080p High with MSAA 2x, which already disproves your claim.

Also:

Away from the general problems have especially older GPUs from AMD are massively struggling with Forza Horizon 3. Polaris provides good results. Pascal Nvidia is the the best choice for Forza Horizon 3

 

[AtenRa]

Some day you ill have to realise that is just a myth, and Nvidia problems on old cards are mostly VRAM limited, not saying its the only issue, but lack of VRAM is playing a mayor role, same thing gona happen to old AMD cards if VRAM usage is forced to go up again, if 4GB is rendered useless for Ultra quality, you can forget about Furys, its gona fell faster than a 2GB GTX770.

GTX 780Ti with 3GB is faster than 4GB R9 290X

GTX 970 with 3.5+0.5GB is faster than 4GB Fury Nano

I will hardly call this a VRAM limitation.

This game needs a patch and/or new driver.