The AMD Mantle Thread

[RampantAndroid]

So what's the debate? It still doesn't show "nVidia doing better than AMD here."

It doesn't disprove it either, if the argument is about nvidia doing better in CPU limited situations, using D3D.

 

[biostud]

Basically Mantle will only help in the situations where the GPU calculations are limited due to waiting on the CPU.

Different scenarios:

1. Oxide demo: Lots of simple drawings, not taxing the GPU that much, but requires lots of CPU draw calls in D3D. Mantle works well in this scenario, since when not limited by the CPU it can render as fast as the GPU is capable of.

2. BF4: Very detailed and GPU taxing. Also very CPU taxing in 64MP. On a modern quad core + video card it runs fine, and Mantle only improves a little. With a slower CPU Mantle can help a great deal. With a modern quad core the GPU is not going to be that limited by the CPU.

3. So with the possibility of increasing the draw calls significantly it will be possible to add lots "simple" objects with only a small performance hit. Something that could be useful for something like total war series or debris from explosions. (or maybe tons of physX objects flying around )

 

[biostud]

Bit-tech found very small gains with Mantle over DX11.
Even when underclocking the CPU down to 2.8GHz

And with R9 280x and down, they sometime got negative results when using Mantle over DX11.

http://www.bit-tech.net/hardware/graphics/2014/02/05/amd-mantle-bf4-performance/

The release notes already states that only 290/290x and 260x are optimized for Mantle at this time. the R9 280X and 7xxx cards will be optimized later.

 

[ASM-coder]

Basically Mantle will only help in the situations where the GPU calculations are limited due to waiting on the CPU.

I think most of us would agree with this and your 3 scenarios.
I suppose the open debate is that NVidia doesn't need Mantle, because
the drivers are so much better. But only at Win 8 and only with the latest
BF4 patch, so nobody has really benchmarked this yet. At least that's
what I think the argument is.

 

[itsmydamnation]

except for where all the data doesn't support that. And gpus performance improvements per gen are far greater then CPU so every gen your pushing more and more cpu bottlenecks. its also ignores all the exposure to the hardware that DX11 limits, Repi has commented that BF4 mantle has additional MSAA optimisations. The fact that mantle is beta, the driver is beta and this is the first release of mantle on BF4.


whats interesting is that there is one outlying site in mantle benchmarking for multiplayer and thats this one http://pclab.pl/art55953-3.html yet yet thats the one people here keep posting.......

edit: u'll notice that the DX11 data aligns across all three, yet mantle doesn't for pclab...........

http://www.golem.de/news/amds-mantle-api-im-test-der-prozessor-katalysator-1402-104261-3.html
http://www.pcgameshardware.de/AMD-R...ecials/AMD-Mantle-Test-Battlefield-4-1107754/

i dont get what people dont want to analyze this properly, they would rather "WIN"..........

 

[ASM-coder]

except for where all the data doesn't support that.

And common sense doesn't support it either. Dice would not have been asking GPU vendors for a better API. Dice wouldn't have invested the time, they would have told AMD to figure out how to fix their drivers. And NVidia would be showing off their own benchmarks.
And lastly, without any evidence to prove it, I can say with confidence that:
Any program(in this case DX) that supports multiple version of multiple products with a wide range of capabilities that have evolved over a number of years, will not be as efficient as it could be.

 

[Paul98]

I did a quick test in BF4 at three resolutions, 1366x768, 1920x1080, and 2560x1440 using the settings I play with. This is using i5 and 7850

In test range
2560x1440 ~Same frame rates
1920x1080 5-10% more fps
1366x768 30-50% more fps
unless stuttering all resolutions get smoother frame rates with mantle.

If I use vsync, or gametime.maxvariablefps GPU usage is much smoother. With mantle I just get a strait line that goes up and down with gpu usage, where is DX I get lots of spikes even though it doesn't spike over the max frame rate.

 

[Carfax83]

I doubt it is related only to a better threading, I would assume it is related to a lower CPU overhead as well, Nvidia seems to handle high drawcall numbers better than AMD under directx 11.

The high CPU overhead primarily comes from the API, or Direct3D. Better threading is one of the best ways to reduce your CPU overhead, as the workload is distributed across more threads.

I'm sure there are other factors involved as well (like the DX11.1 constant buffers CPU optimization), but this is the primary one I believe. I base this on the PClabs.pl review which clearly shows that AMD drivers have much poorer scaling past two threads, whereas NVidia's drivers easily scale up to four and beyond, provided the game supports that many threads..

 

[Spjut]

I did a quick test in BF4 at three resolutions, 1366x768, 1920x1080, and 2560x1440 using the settings I play with. This is using i5 and 7850

In test range
2560x1440 ~Same frame rates
1920x1080 5-10% more fps
1366x768 30-50% more fps
unless stuttering all resolutions get smoother frame rates with mantle.

If I use vsync, or gametime.maxvariablefps GPU usage is much smoother. With mantle I just get a strait line that goes up and down with gpu usage, where is DX I get lots of spikes even though it doesn't spike over the max frame rate.

Was that single or multiplayer?

PCgameshardware's test shows too that Mantle is great for users with older CPUs. Mantle doesn't come fully to its right when most reviewers still are using relatively high-end CPUs in their benchmarks

Mantle isn't that important for gamers that upgrade once in every two years, which I'd guess most on Anandtech do, but it's great for those with older PCs

 

[jj109]

So what's the debate? It still doesn't show "nVidia doing better than AMD here."

If you'd read the original post, the poster was trying to use it as a CPU-bound example where AMD's DX11 drivers were keeping up with Nvidia. However, it is not a CPU bound example. Not even close.

 

[itsmydamnation]

The high CPU overhead primarily comes from the API, or Direct3D. Better threading is one of the best ways to reduce your CPU overhead, as the workload is distributed across more threads.

I'm sure there are other factors involved as well (like the DX11.1 constant buffers CPU optimization), but this is the primary one I believe. I base this on the PClabs.pl review which clearly shows that AMD drivers have much poorer scaling past two threads,

it is interesting data

whereas NVidia's drivers easily scale up to four and beyond, provided the game supports that many threads..

from the data the game has little to do with it, even SC2 sees core scaling on NV thats a DX9 game. so the question is whats NV doing differently across the board in the way the drivers operate. You can see at low thread counts that they get poor clock scaling, if i had to guess i would say that's synchronization costs across the threads/cores.

but when you compare this to there BF4 benchmark it still doesn't align, as it has NV scaling with CPU clock on 4770k, AMD not scaling with clocks in DX and there mantel data appears flat out wrong. the I3/pentium data for there BF4 review also doesn't align tot he trends in this review.

I think there is enough variability there for anyone to see what they want to see.

 

[krumme]

We have to remember Cpu frame to frame is more inconsistent than gpu frame to frame.
What it means imho is if your api doesnt smooth out the cpu frame to frame or the cpu is just flat out a good deal faster, the frames you will see on your screen will jitter although the average fps is okey.

Thats how i understand it and thats why consoles have this fluent frame to frame look, and what mantle aparently also have.

 

[Carfax83]

no you haven't and no it isn't... i notice that you haven't explained the bit i pointed out as well.

the funny thing is lots of people who work with or on GPU's say DX multithreaded submission say it is fundamentally broken.

Who on Earth said anything about DX multithreaded submission? I was talking about basic driver multithreaded enhancements, something which NVidia has been doing for almost 15 years with their drivers when the first dual core processors became available:

NVidia drivers multithreaded since 2005

But if DX11 multithreading is broken, how on Earth did NVidia manage to squeeze out a 50% gain (in CPU limited situations) over AMD in Civilization 5?

And why is Slightly Mad Studios, the developers of Project CARS using the technology in their game?

In fact, in this comparison posted video posted on youtube last year pitting a GTX 660 vs a HD 7850, the GTX 660 has a 55% lead over the AMD card, courtesy of DX11 multithreading:

GTX 660 vs HD 7850

So apparently, DX11 multithreading does work, at least on NVidia hardware.

and here more of the stuff im talking about, possible reasons for NV's overall superior showing.

That's just a bunch of speculation. The PClabs.pl review shows that AMD's drivers have poor scaling on multicore processors.

I'll take that over some random internet post on the net any day.

Drivers can get in the way of performance but they dont magically create higher then peak performance as determined by the hardware. The simple fact mantle scales so well on low end CPU's when NV doesn't ( magical DX multithreading working so well!) yet scale CPU performance high enough they end up in front and AMD stop scaling at not much more then a 7850K.

Mantle is unquestionably more efficient than Direct3D, but it's not huge unless we're comparing AMD to AMD.

The 7850K plus 780 Ti pairing was only 18% slower than the 7850K plus R290x pairing running on Mantle.

But when you compare the 7850K plus R290x pairing running on DX to Mantle, Mantle gives you almost 60% more performance..

The problem basically, is that AMD's DX drivers suck ass..

There's no question that NV seems CPU limited up to around a 4.5ghz 4770K in multiplayer. but thats not the question, the question is why do they have so much more performance when not CPU limited vs Mantle on hardware that on many other games is far closer. ie 290x battling the 780 and not the 290.

Where are you getting "NV seems CPU limited up to around a 4.5ghz 4770K in multiplayer."

I posted some benchmarks many pages back comparing my 4.5ghz 3930K vs stock clocks at the South China Sea mission in the SP campaign, the same area AMD used to showcase their Mantle improvements as it's very CPU intensive.

Overclocking the CPU surprisingly had little performance gain, around 10 FPS or so, which considering I was already in triple digit frame rates, wasn't much for a 1ghz overclock.

So on my setup, I wasn't really CPU limited at all; at least not appreciably so. Frostbite 3 engine can scale up to 8 threads, so as long as you have a strong multithreaded processor, you're unlikely to be CPU limited to the point where it affects gameplay.

also if DX multithreading works so well, why does it gen pwnd so hard by openGL extensions?
https://static.slo-tech.com/52734.jpg
https://static.slo-tech.com/52736.jpg

You post questionable benchmarks from God knows how many years ago, and expect them to have any sort of impact today?

DX11 multithreading performance is wholly contingent on the drivers. From the time when NVidia first introduced DX11 multithreading capable drivers with the 270xx, to the very recent 334.67 drivers, the performance has undoubtedly increased.

So that's not really a good point.

 

[krumme]

Interesting interview just posted on PCper. It covers some of the "how hard will it be to port DX to Mantle and vice versa" type questions.

http://www.pcper.com/reviews/General-Tech/Core-Mantle-Questions-Interview-AMDs-Guennadi-Riguer

Thanx for link. I notice this about the compute perspective:
"
Q: When I see Mantle, it feels like AMD is making a play for compute-based game engines. Specifically, I see the minimizing of strict shaders (one which runs per vertex, one which runs per pixel, etc.) and the emphasis of keeping shader units constantly loaded with relevant math. What future do you see for the graphics pipeline as a whole? How about fixed-function features?

[Guennadi] Pipeline concept in Mantle provides a nice abstraction that is forward looking and allows to move graphics further into the compute shader territory. It has to be said: except for a few slight differences in terms of how compute and graphics shaders are fed data and how they tie into the flow of data in the pipeline, they are very similar and are running on the same shader core. Right now the graphics pipeline is fairly rigid in terms of the data flow, but hopefully in the future it will become more flexible. The fixed-function parts won't completely go away since they provide very high power and performance efficiency, but they could evolve to support more flexible pipeline configurations. This is a pipe dream at this point, pun intended J, but in the future it can definitely be a reality. Definitely, more research is needed in this direction.
"

 

[Skurge]

Who on Earth said anything about DX multithreaded submission? I was talking about basic driver multithreaded enhancements, something which NVidia has been doing for almost 15 years with their drivers when the first dual core processors became available:

NVidia drivers multithreaded since 2005

But if DX11 multithreading is broken, how on Earth did NVidia manage to squeeze out a 50% gain (in CPU limited situations) over AMD in Civilization 5?

And why is Slightly Mad Studios, the developers of Project CARS using the technology in their game?

In fact, in this comparison posted video posted on youtube last year pitting a GTX 660 vs a HD 7850, the GTX 660 has a 55% lead over the AMD card, courtesy of DX11 multithreading:

GTX 660 vs HD 7850

So apparently, DX11 multithreading does work, at least on NVidia hardware.

That's just a bunch of speculation. The PClabs.pl review shows that AMD's drivers have poor scaling on multicore processors.

I'll take that over some random internet post on the net any day.

Mantle is unquestionably more efficient than Direct3D, but it's not huge unless we're comparing AMD to AMD.

The 7850K plus 780 Ti pairing was only 18% slower than the 7850K plus R290x pairing running on Mantle.

But when you compare the 7850K plus R290x pairing running on DX to Mantle, Mantle gives you almost 60% more performance..

The problem basically, is that AMD's DX drivers suck ass..

Where are you getting "NV seems CPU limited up to around a 4.5ghz 4770K in multiplayer."

I posted some benchmarks many pages back comparing my 4.5ghz 3930K vs stock clocks at the South China Sea mission in the SP campaign, the same area AMD used to showcase their Mantle improvements as it's very CPU intensive.

Overclocking the CPU surprisingly had little performance gain, around 10 FPS or so, which considering I was already in triple digit frame rates, wasn't much for a 1ghz overclock.

So on my setup, I wasn't really CPU limited at all; at least not appreciably so. Frostbite 3 engine can scale up to 8 threads, so as long as you have a strong multithreaded processor, you're unlikely to be CPU limited to the point where it affects gameplay.

You post questionable benchmarks from God knows how many years ago, and expect them to have any sort of impact today?

DX11 multithreading performance is wholly contingent on the drivers. From the time when NVidia first introduced DX11 multithreading capable drivers with the 270xx, to the very recent 334.67 drivers, the performance has undoubtedly increased.

So that's not really a good point.

CIV5 doesn't have a 50% lead for nvidia anymore. I play project cars a lot and I switch GPUs very often. Yes the game runs much faster on Nvidia hardware without even looking at the framerate. That is not because of the CPU, that cause the AMD drivers are not optimised for this ALPHA game. I know it isn't the CPU cause the things that hit amd cards the most is AA mostly and that is in a very GPU limited scenario. Project cars is very GPU limited. I can tell you right now CPU makes very little difference in that game regardless of the GPU.

Even in your own video, you can see the settings are maxed out. If it were indeed cause of CPU limitation the higher the settings the smaller the gap. There is a 7850 at work and I am currently running a GTX660. I have both Project cars and CIV5, I also have BF4 I will do some benchmarks for you for 3 games this weekend with my system at various resolutions and CPU clocks/core in DX and mantle.

 

[ASM-coder]

But if DX11 multithreading is broken, how on Earth did NVidia manage to squeeze out a 50% gain (in CPU limited situations) over AMD in Civilization 5?

Old Anandtech review.

http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/16
Remember when NVIDIA used to sweep AMD in Civ V? Times have certainly changed in the last year, that’s for sure. It only seems appropriate that we’re ending on what’s largely a tie. At 2560 the GTX 680 does have a 4% lead over the 7970, however the 7970 reclaims it’s lead at the last possible moment at 1920. At this point we’ve seen the full spectrum of results, from the GTX 680 losing badly to winning handily, and everything in between.
http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/17
"Remember when NVIDIA used to sweep AMD in Civ V Compute? Times have certainly changed. AMD’s shift to GCN has rocketed them to the top of our Civ V Compute benchmark."

 

[Carfax83]

CIV5 doesn't have a 50% lead for nvidia anymore.

The very last time Anandtech benched Civ 5 for a GPU review (the GTX 770), here were the results:

The GTX 780 is almost 50% faster than the 7970 GE, and these are at GPU LIMITED SETTINGS. If they had more CPU limited settings, then the results would likely have been higher, as DX11 multithreading is designed to get rid of CPU bottlenecks.

Why am I using the GTX 780 instead of the GTX 770 you might ask? It's simple. The GTX 780 is the most CPU limited card in the benchmark.

In fact, the benchmark had to be retired as modern cards are too powerful and are maxing it out.

In the original review where NVidia first introduced driver command lists in their 270 driver:

Nearly 55% faster.

I play project cars a lot and I switch GPUs very often. Yes the game runs much faster on Nvidia hardware without even looking at the framerate. That is not because of the CPU, that cause the AMD drivers are not optimised for this ALPHA game.

Lack of optimization is playing a role, but so is DX11 multithreading, which the engine supports. We're talking 55% here, that's huge, and mirrors past comparisons between NVidia and AMD in Civilization 5.

I know it isn't the CPU cause the things that hit amd cards the most is AA mostly and that is in a very GPU limited scenario. Project cars is very GPU limited. I can tell you right now CPU makes very little difference in that game regardless of the GPU.

We don't know what settings the guy used, but I have a hard time believing a racing game (particularly one like this) isn't going to be CPU limited, as you're moving so fast so things have to be rendered quickly.

The CPU is part of the rendering process. The GPU can't render nothing without the CPU giving the command. If the CPU isn't sending info to the GPU fast enough, you run into a bottleneck.

Even in your own video, you can see the settings are maxed out. If it were indeed cause of CPU limitation the higher the settings the smaller the gap. There is a 7850 at work and I am currently running a GTX660. I have both Project cars and CIV5, I also have BF4 I will do some benchmarks for you for 3 games this weekend with my system at various resolutions and CPU clocks/core in DX and mantle.

The IQ settings may have been high, but the resolution was 1280x720, which effectively makes it CPU limited.

 

[Carfax83]

Old Anandtech review.

http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/16
Remember when NVIDIA used to sweep AMD in Civ V? Times have certainly changed in the last year, that’s for sure. It only seems appropriate that we’re ending on what’s largely a tie. At 2560 the GTX 680 does have a 4% lead over the 7970, however the 7970 reclaims it’s lead at the last possible moment at 1920. At this point we’ve seen the full spectrum of results, from the GTX 680 losing badly to winning handily, and everything in between.
http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/17
"Remember when NVIDIA used to sweep AMD in Civ V Compute? Times have certainly changed. AMD’s shift to GCN has rocketed them to the top of our Civ V Compute benchmark."

You posted the COMPUTE benchmarks, not the actual game benchmarks.

I posted the latest Civilization 5 game benchmark on Anandtech above.. The GTX 680 and 770 are both more GPU limited at those settings, than the GTX 780..

 

[el etro]

For the Mantle question:

Battlefield 3: Game launched, Fermi vs Northern islands, Multiplayer.

Battlefield 3: Same test, Kelper vs GCN, one and half year later:

 

[Hitman928]

Carfax, your argument for Civ5 makes no sense to me. Besides that, I found a test with an x290 and a 780 in Civ5.

http://www.overclockers.com/amd-r9-290-graphics-card-review

In an even more cpu stressful situation the Nvidia cards still don't show that much of a lead. They are leading the AMD cards, sure, but not by anything you'd find too unusual outside of game by game variance. I think you have a point with the Nvidia drivers being more efficienct than the AMD ones in general, but not anywhere near the level you are trying to make it out to be (or the effect of dx11 multithreading).

 

[ASM-coder]

You posted the COMPUTE benchmarks, not the actual game benchmarks.

I meant to post both. I thought I could just change the 17 to 16, but both went to the same page.
http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/16

 

[Shivansps]

Keep in mind that CIV5 will give you way different results depending on the testing. Its not the same a huge map with lots of units on visual than just starting a game in a tiny map, it happens the same thing than in BF4.

 

[DarkKnightDude]

Doesn't Civ 5 have a dedicated benchmarking tool?

 

[bystander36]

Doesn't Civ 5 have a dedicated benchmarking tool?

Many review sites, if not most, do not use the dedicated benchmarking tools supplied by games. Unless specified, you have no idea what is used. The better sites tend to tell you how it was tested, and it surprised me how often they do not use supplied benchmark tools.

 

[GaiaHunter]

Why am I using the GTX 780 instead of the GTX 770 you might ask? It's simple. The GTX 780 is the most CPU limited card in the benchmark.

So what you are saying is that for comparable graphic cards from AMD and NVIDIA there is no difference.

I'm not sure how that is proving your point.