AMD Mantle API [updating]

Another possible data point in the discussion:

http://www.youtube.com/watch?feature=player_detailpage&v=JTWkpnbdROg#t=174

Note the variation in CPU load for the DX case and compare it to Mantle.

It looks like DX causes greater fluctuations.

Whether or not this affects apparent smoothness is another question.
If it is an additional source of fluctuations in the time differences between the snapshots of the game world and the finishing of the rendering/dipslay on the monitor, it likely does.
 
Last edited by a moderator:
Yeah on AMD the frame consistency is definitely improved with Mantle, but the interesting result from the TR review is that NVIDIA's DX driver is much better than AMDs in this regard (at least in BF4), and competitive with the Mantle result in smoothness.

Don't want to overstate it though, as the variations are fairly small. It's the spikes that are more concerning, and Mantle doesn't yet seem to do away with those (they may be purely in the game/engine, it's impossible to know from this data).
 
Don't want to overstate it though, as the variations are fairly small. It's the spikes that are more concerning, and Mantle doesn't yet seem to do away with those (they may be purely in the game/engine, it's impossible to know from this data).

I wonder if maybe it's Windows's scheduler that is the problem.

Since Mantle tries to use every available core/thread for rendering, any new process that needs CPU is going steal a bit from a thread possibly involved in rendering. With DX, there's usually a spare core or thread available for these interruptions. What happens when Windows has to rejuggle all those live threads? The Mantle threads/GPU seem to pause and do nothing for a moment. Perhaps Windows is taking its time to transfer control back to the rendering threads.

The spikes see very periodic, as if some background process is waking up every second or so.
 
My experience (290CF) with the 14.1 drivers...

BSOD after BSOD²....BSOD*10²²
And I'm not even talking about Mantle, but every dx10 and dx11 game.
Also dx9 disables CF for now.

It's nice they develop this CPU-friendly route, but it seems it comes at the cost of resources for DX.
The experience I have with 290CF is mediocore at the very best.
In november cat. 13.11beta V9.4 starts to introduce a bug in conjunction with v-sync where the perf. plummets and the audio get distorted.
 
I wonder if maybe it's Windows's scheduler that is the problem.
Always possible, but you're really just talking about the difference of being truly "CPU bound" on every thread vs. bound on a single thread with idle cores. It's not clear to me that even Mantle gets to the former case on quad-core CPUs, but maybe someone has tested that? In any case it would still be surprising to see spikes of the magnitude shown here just due to some thread getting descheduled briefly, and it wouldn't be a situation that only affects Mantle.

BSOD after BSOD²....BSOD*10²²
And I'm not even talking about Mantle, but every dx10 and dx11 game.
Hmm sounds like like the driver might have made non-trivial changes to the kernel mode driver (the UMD can't blue-screen the system). Brings us back to the point of wondering what exactly Mantle/new driver is doing to interact with WDDM... my guess is some of the changes would have to be related to the GPU page table stuff in Mantle, but it's hard to know without better information on exactly how Mantle handles stuff like WDDM deciding to move pages/VAs around, etc.
 
Yeah on AMD the frame consistency is definitely improved with Mantle, but the interesting result from the TR review is that NVIDIA's DX driver is much better than AMDs in this regard (at least in BF4), and competitive with the Mantle result in smoothness.
At those particular graphics settings.

I'm still waiting for evidence across a range of graphics settings.
 
At those particular graphics settings.
The settings were pretty much the highest IIRC... graphics settings other than shadows/view distance usually don't have a significant effect on CPU time. The more data the better but do you have a strong reason to believe it will vary much?

The higher the performance gets (say on low settings) the less interesting it is as well. TR's 99% metric is really a pretty good "single number"... variance between 1-2ms is not really an issue for instance but 16-30 is a huge issue (and depending on the frequency should be considered as running at the slower speed, hence the metric).
 
The settings were pretty much the highest IIRC... graphics settings other than shadows/view distance usually don't have a significant effect on CPU time. The more data the better but do you have a strong reason to believe it will vary much?
The Tech Report article used "high" settings rather than "Ultra" as he was trying to induce some extra CPU limitations, however Mantle wasn't scaling as much on the higher CPU's which indicates that while there is more CPU limitation than the highest graphical settings it wasn't completely CPU bound as you see Mantle scale much more in those cases.
 

it's a typical game in terms of CPU usage, BF4 is more unusual, that's why Intel CPUs are so good at gaming

ac4bf_cpu_amd.png


ac4bf_cpu_nv.png


but look at the Phenom II X4 and G3220 (2c)

Nvidia drivers suffer more with 2 cores (or should I say "threads" because the i3 with HT works great), but have good scaling with more cores (threads), while AMD works great with 2 cores but scales less with more cores!?

the test from the link you posted was made using AMD graphics.
 
anyone else drooling to see a mantle build of civ 5 ? Late in the game my cpu (bulldozer fx 8150 ) is dead trying to keep up with things on larger maps with tons of city states and cities .
 
The Tech Report article used "high" settings rather than "Ultra" as he was trying to induce some extra CPU limitations, however Mantle wasn't scaling as much on the higher CPU's which indicates that while there is more CPU limitation than the highest graphical settings it wasn't completely CPU bound as you see Mantle scale much more in those cases.
Absolutely, I'm just not sure why Jawed thinks the variance would be different on ultra. I doubt that introduces much more CPU load, and deferred MSAA tends to introduce a pile more GPU load.
 
There's a forum on AMD site for driver / hardware related problem, now i think they count too much on the community. http://forums.amd.com/game/

Yep that's the one I was referring to. I agree with you.

That place seem to have a strange aura around it. It must be, but not limited to feng shui. AMD needs to re-architecture their forum like they did with GCN.

but look at the Phenom II X4 and G3220 (2c)

Nvidia drivers suffer more with 2 cores (or should I say "threads" because the i3 with HT works great), but have good scaling with more cores (threads), while AMD works great with 2 cores but scales less with more cores!?

the test from the link you posted was made using AMD graphics.

Great find. AC4 is a NVIDIA sponsored-game (GameWorks title?), maybe it also has something to do with it.
 
Absolutely, I'm just not sure why Jawed thinks the variance would be different on ultra. I doubt that introduces much more CPU load, and deferred MSAA tends to introduce a pile more GPU load.
I'm curious to see evidence. Both at higher and lower graphics settings. A single data point is only mildly interesting.

I expect more advanced graphics settings to increase draw calls per frame. There's too many variables in play to theorise on the net effect on absolute frame rate.

And I'm curious to see if specific types of graphics settings favour one GPU architecture/driver.

It's tempting to say that AMD needs Mantle because D3D is such a bad match for its GPU/driver architecture. But I think a bit more evidence is required.
 
I'm curious to see evidence. Both at higher and lower graphics settings. A single data point is only mildly interesting.

I expect more advanced graphics settings to increase draw calls per frame. There's too many variables in play to theorise on the net effect on absolute frame rate.

And I'm curious to see if specific types of graphics settings favour one GPU architecture/driver.

It's tempting to say that AMD needs Mantle because D3D is such a bad match for its GPU/driver architecture. But I think a bit more evidence is required.

Perhaps the best tests should involve complicated scenes with many textures and small objects. Not CPU bound, exactly, but call/second bound.

The Star Swarm demo seems to be one of the few examples of such. If Civ5 gets mantle, that might also be a good example.
 
It's tempting to say that AMD needs Mantle because D3D is such a bad match for its GPU/driver architecture. But I think a bit more evidence is required.
I don't think I'd go that far even with "more evidence", but I also don't think it's unreasonable to conclude that NVIDIA's army of software folks earns their pay here.

And again, obviously none of this is to say that Mantle doesn't reduce overhead, etc. Clearly it does (see Star Swarm, etc), but I still reject the notion that reducing any harmful frame variance in BF4 requires Mantle. I think NVIDIA has a fairly clear counter-example to that, and even a single data point throws that claim into question.
 
Back
Top