Sniper Elite 4: Performance Analysis

[W1zzard]

We test Rebellion's new shooter on 10 modern graphics cards, with the latest game-optimized drivers from AMD and NVIDIA. A surprise is that AMD gains up to 27% performance from switching to DX12 with Async Compute, while the performance uplift for NVIDIA users is only up to 5%.

Show full review

 

[EzioAs]

Probably not a very popular opinion but if you tweak the settings a little bit, I'm thinking the GTX 1060 and RX 480 would be more than adequate for 1440p 60fps (avg). Some people seem to think those 2 GPUs are exclusively for res up to 1080p but I very much disagree on that. It really boils down on what the user wants in the end.

NOTE: Not that I'm saying you're wrong in your conclusion, @W1zzard. You clearly state that the GPUs I mentioned are for THE highest detail settings at 1080p.

 

[Fluffmeister]

With the gains AMD show on DX12, i assume the DX11 performance graphs were to horrible to show?

 

[Xzibit]

Either there is a typo on the RX 480 FPS # or it should switch places with the GTX 1060

 

[W1zzard]

Either there is a typo on the RX 480 FPS # or it should switch places with the GTX 1060

Oh, I manually sort them and missed that one. Will fix tomorrow morning

 

[VSG]

Do you happen to know what the standard deviation was for the FPS results? Perhaps not for all the cards and all resolutions, but some idea would tell a nice story.

 

[Fragment]

Hi!

Nice review, although... what are the clock speed readouts of the cards during the run?
You write the CPU clock but nowhere the GPU clock is mentioned, which is quite important to know in a graphics benchmark. Even if they are stock, it would be important to know WHICH stock clock it is.

Thanks!

 

[Deleted member 50521]

The only exception is the R9 Fury X which does extremely well, beating the GTX 1070 at 4K, sitting right between GTX 1070 and 1080. What makes this victory even more memorable is that Fury X "only" has 4 GB of VRAM. At least 4 GB of VRAM seems to be the ideal memory configuration for Sniper Elite 4, to run at highest details. Our memory usage testing reveals that the game doesn't use excess memory and today's gaming cards won't run into trouble, memory capacity-wise.

I am soooooooo happy that I bought my lovely FuryX. Hopefully with more Aysnc Compute enabled titles the full 4096SP will be utilized better in the future.

 

[Deleted member 50521]

Also an interesting note regarding HBM VRAM. In recent AMD VEGA events it seems they have stopped calling HBM as VRAM, but rather HBC=High Bandwidth Cache. So I can only assume somewhere down the pipeline AMD will have a card that has ~8GB of HBM as Cache and >16GB of GDDR5/GDDR5X/GDDR6 as VRAM. That would be very interesting.

 

[renz496]

Also an interesting note regarding HBM VRAM. In recent AMD VEGA events it seems they have stopped calling HBM as VRAM, but rather HBC=High Bandwidth Cache. So I can only assume somewhere down the pipeline AMD will have a card that has ~8GB of HBM as Cache and >16GB of GDDR5/GDDR5X/GDDR6 as VRAM. That would be very interesting.

will they mixed the usage of HBM and GDDR memory? isn't that going to beat the purpose of going HBM in the first place? AFAIK what AMD intend to do is something different. high capacity HBM is very expensive. so they want games to reduce VRAM usage with cleaver usage of HBC. but we don't know how transparent is this to game and operating system. do we need game developer specifically to code for it? does it need to be included in 3D API spec? just look at async itself. AMD have the hardware since the very first GCN but they end up leaving it idle in their hardware because DX11 cannot access the feature.

 

[londiste]

With the gains AMD show on DX12, i assume the DX11 performance graphs were to horrible to show?

doesn't seem to be too bad:
http://gamegpu.com/action-/-fps-/-tps/sniper-elite-4-test-gpu

 

[W1zzard]

Hi!

Nice review, although... what are the clock speed readouts of the cards during the run?
You write the CPU clock but nowhere the GPU clock is mentioned, which is quite important to know in a graphics benchmark. Even if they are stock, it would be important to know WHICH stock clock it is.

Thanks!

we've always used reference design cards at reference clocks for all GPU related articles

 

[londiste]

with pascal and gpu boost 3.0, reference clocks don't necessarily mean all that much

 

[pat-roner]

with pascal and gpu boost 3.0, reference clocks don't necessarily mean all that much

I would guess that most people buy cards and run them stock. A fair deal of people are also buying reference design

 

[londiste]

what i mean was that gpu boost 3.0 lets the clock increase a lot provided that power or tdp limit is not reached.

for example, gtx1080 reference clock is 1607mhz and boost clock is 1733mhz.
- founders edition gtx1080 cards (depending on temperature inside the case) generally average to little above the boost clock .
- aib gtx1080 with openair coolers (again, depending on case airflow, even more so than fe cards) will usually average considerably higher.

specific example - my gainward phoenix with its annoyingly huge 2.5slot cooler averages to around 1850mhz inside somewhat warm case and stays at 1885mhz outside a case. slapping a fullcover waterblock on the same card (with max 40-45c temperatures at full load) results in clock being stuck at 1885mhz at load. 1885mhz seems to be the highest bin for boost in this case.

all of this is at reference clocks, with cooling being the only difference.
tpu review of fe gtx 1080 had average clock of 1785mhz. 1885 is 5.6% over that and 8.7% over the 1733mhz boost clock. these are not negligible differences.

this behaviour is currently unique to pascal and gpu boost 3.0. yes, previous gen cards and amd cards have varying degrees of boost for clocks but not to this degree.

 

[Fluffmeister]

http://www.pcgameshardware.de/Sniper-Elite-4-Spiel-56795/Tests/Direct-X-12-Benchmark-1220545/

Seems even Kepler gains going to DX12.

 

[rtwjunkie]

what i mean was that gpu boost 3.0 lets the clock increase a lot provided that power or tdp limit is not reached.

for example, gtx1080 reference clock is 1607mhz and boost clock is 1733mhz.
- founders edition gtx1080 cards (depending on temperature inside the case) generally average to little above the boost clock .
- aib gtx1080 with openair coolers (again, depending on case airflow, even more so than fe cards) will usually average considerably higher.

specific example - my gainward phoenix with its annoyingly huge 2.5slot cooler averages to around 1850mhz inside somewhat warm case and stays at 1885mhz outside a case. slapping a fullcover waterblock on the same card (with max 40-45c temperatures at full load) results in clock being stuck at 1885mhz at load. 1885mhz seems to be the highest bin for boost in this case.

all of this is at reference clocks, with cooling being the only difference.
tpu review of fe gtx 1080 had average clock of 1785mhz. 1885 is 5.6% over that and 8.7% over the 1733mhz boost clock. these are not negligible differences.

this behaviour is currently unique to pascal and gpu boost 3.0. yes, previous gen cards and amd cards have varying degrees of boost for clocks but not to this degree.

You got the reference clocks part right and focused on that, while mentioning different coolers. Apparently you missed the part where W1zzard also said reference design, which in the case of Pascal would be FE models.

 

[Kaotik]

Also an interesting note regarding HBM VRAM. In recent AMD VEGA events it seems they have stopped calling HBM as VRAM, but rather HBC=High Bandwidth Cache. So I can only assume somewhere down the pipeline AMD will have a card that has ~8GB of HBM as Cache and >16GB of GDDR5/GDDR5X/GDDR6 as VRAM. That would be very interesting.

Nope.
AMD's plan, if it works, is to get only the data you need on the video card, rather than all the data the game wants to allocate (there can gigabytes of difference on allocated mem and actually used mem). HBM2's role in this scenario is to hold the data that's actually being accessed, the rest should reside on your system RAM, NVRAM, SSD, HDD, heck, even network storage is supported. Future Radeon Pro SSG -models will probably benefit from this even more, as they house SSDs of their own. Vega supports 512TB virtual address space to accomodate all this.

 

[Basard]

http://www.pcgameshardware.de/Sniper-Elite-4-Spiel-56795/Tests/Direct-X-12-Benchmark-1220545/

Seems even Kepler gains going to DX12.

Maybe something to do with DX12 using more cores in the CPU.... I dunno, I'm not an expert.

 

[londiste]

You got the reference clocks part right and focused on that, while mentioning different coolers. Apparently you missed the part where W1zzard also said reference design, which in the case of Pascal would be FE models.

my bad, you're probably right.

Maybe something to do with DX12 using more cores in the CPU.... I dunno, I'm not an expert.

dx12 is much more than just async compute.
that german article actually shows that very clearly with turning on async compute giving only a small performance boost even on usual suspects for good boost (furyx, 390x) and dx12 in general giving a much larger boost on all cards.

 

[FYFI13]

doesn't seem to be too bad:
http://gamegpu.com/action-/-fps-/-tps/sniper-elite-4-test-gpu

Doooh, this game looks like one of Call of duty tittles back in 2006, yet needs better than average rig to run it on max details. I'll pass.

 

[rtwjunkie]

Doooh, this game looks like one of Call of duty tittles back in 2006, yet needs better than average rig to run it on max details. I'll pass.

Oh no, it's extremely good looking! I actually was not prepared for the increase of visual details and fidelity in the game beyond where it was for SE3.

Also, the maps are huge, the AI is a lot better, and I can go almost anywhere I want. Excellent for finding those perfect sniper nests!

 

[HD64G]

Great implementation of DX12 that helps FuryX to show off its power. Being close to 1080 on 4K with half its VRAM and almost half its clock speed is great imho and shows that CGN is future proof as much as it gets. Vega will probably be too good to be true on DX12 and Vulcan.

 

[TheinsanegamerN]

Great implementation of DX12 that helps FuryX to show off its power. Being close to 1080 on 4K with half its VRAM and almost half its clock speed is great imho and shows that CGN is future proof as much as it gets. Vega will probably be too good to be true on DX12 and Vulcan.

While i agree, lets wait for AMD to actually deliver VEGA first before getting all excited about it.

 

[Batou1986]

Where are the DX11 benchmark numbers tho without them the chart is meaningless.
If AMD cards pulling 40fps in DX11 and 60fps in DX12 but Nvidia was getting 55fps in DX11 and 60fps in DX12 makes that 27% moot.