Mantle API presentation by AMD, DICE and Oxide - AMD Summit 2013

[Steevo]

Multithread the AI and then intermix units 50 across that unit type and allow use of one call for all of them, 50K render units just dropped to 1K actually rendered in animation and an equal number of setup calls, improving efficiency.

 

[crazyeyesreaper]

Yea something like that. Add to that physics calculations as in Total War physics for charges is calculated on a unit / soldiers left in unit. Overall the CPU limitation could with Mantle be removed for the most part. Sadly That would require CA re writting their game engine and support from AMD with mantle. Considering CA wasted 40% of the Rome 2 budget on a new building and the Game itself was the worst out of the entire series. Its doubtful those key changes will happen. Ah well another one of those OH THE POSSIBILITIES of Mantle moments. That wont come about for years.

 

[Mathragh]

Fudzilla article about AMD conference yesterday.
Yeah I know its Fudzilla, but they claim the slides from the conference said the BF4 patch would improve FPS by up to 45% on kaveri.

Perhaps the most interesting part of the presentation involved Mantle. AMD claims it will deliver a performance boost of up to 45 percent in Battlefield 4 and in some cases we could be looking at even more

Do with it what you will!

Personally, if they say up to 45%, it'll realistically probably be closer to 30%, which in the end still not that bad, certainly if the min fps spikes also improve as a result of less bottlenecking.

Edit: fixed link

 

[RoutedScripter]

Oh yeah

 

[xenocide]

I don't buy those for a second. In a "Very CPU Intensive" Game an i7-4700K will walk all over any APU, definitely more than a 4fps difference to be had there. Looking at old benchmarks the current line of AMD APU's are still about half the performance of Intel's higher end Desktop CPU's with identical graphics configurations. This is implying that from the current line of AMD APU's (Richland) to Kevari we're to expect a 80% CPU performance gain from a hardware revision alone? Unlikely...

 

[The Von Matrices]

I don't buy those for a second. In a "Very CPU Intensive" Game an i7-4700K will walk all over any APU, definitely more than a 4fps difference to be had there...

I agree. It seems odd that the 4770K is only 50% faster than the Kaveri results in DirectX. I've give the benefit of the doubt and say that the game code can't make use of a quad core+ HT system over a dual module system, which is why the results are closer than if you were to compare a highly threaded process like video encoding, for example.

The GPU in the systems isn't stated either. It could just be something on the low end like a R7 250 and be completely bottlenecked by the GPU.

 

[Mathragh]

I agree. It seems odd that the 4770K is only 50% faster than the Kaveri results in DirectX. I've give the benefit of the doubt and say that the game code can't make use of a quad core+ HT system over a dual module system, which is why the results are closer than if you were to compare a highly threaded process like video encoding, for example.

The GPU in the systems isn't stated either. It could just be something on the low end like a R7 250 and be completely bottlenecked by the GPU.

This would appear to be exactly what oxide developers were "moaning" about when it comes to directx, eventhough dx11 can utilize multiple threads, theres still only one main "game engine" thread which still bottlenecks by loads,

 

[xenocide]

This would appear to be exactly what oxide developers were "moaning" about when it comes to directx, eventhough dx11 can utilize multiple threads, theres still only one main "game engine" thread which still bottlenecks by loads,

But they also claimed that super underclocked AMD CPU's could run programs at the same level as high-end Intel CPU's. If that were the case the Mantle graphs should be identical since it largely removes the CPU bottleneck they claimed to experience--yet the difference between the AMD APU and Intel CPU is nearly identical.

I want to see that Slide 2.7 configuration to see what their setups were because the raw numbers don't add up.

 

[the54thvoid]

I agree. It seems odd that the 4770K is only 50% faster than the Kaveri results in DirectX. I've give the benefit of the doubt and say that the game code can't make use of a quad core+ HT system over a dual module system, which is why the results are closer than if you were to compare a highly threaded process like video encoding, for example.

The GPU in the systems isn't stated either. It could just be something on the low end like a R7 250 and be completely bottlenecked by the GPU.

I think that's the point. It does say "Very CPU limited". That term is generally taken to mean very poor multi core performance. They'll always use worse case scenario's of CPU use to demo oxide running with mantle. Remember that AMD pay for Mantle coding in games so much like other proprietary stuff you'll always get pre-release info that is skewed to favour the thing it is touting.

 

[kn00tcn]

maybe they're doing a single thread, before optimizations? it's a prototype afterall

how do those cpus compare in that sense? still.. 12 to 32fps... very interesting

EDIT:

But they also claimed that super underclocked AMD CPU's could run programs at the same level as high-end Intel CPU's. If that were the case the Mantle graphs should be identical since it largely removes the CPU bottleneck they claimed to experience--yet the difference between the AMD APU and Intel CPU is nearly identical.

that difference must be the AI & everything unrelated to graphics calls

 

[Mathragh]

But they also claimed that super underclocked AMD CPU's could run programs at the same level as high-end Intel CPU's. If that were the case the Mantle graphs should be identical since it largely removes the CPU bottleneck they claimed to experience--yet the difference between the AMD APU and Intel CPU is nearly identical.

I want to see that Slide 2.7 configuration to see what their setups were because the raw numbers don't add up.

Edit: read your text wrong.

I agree, i suppose they really mean that its very cpu limited then, with more than a 100k draw calls(their showcase topped out at 50k i think).

 

[RoutedScripter]

 

[xenocide]

maybe they're doing a single thread, before optimizations? it's a prototype afterall

how do those cpus compare in that sense?

In CPU limited situations under DirectX the 4770K performs about twice as well as the highest end Richland CPU (there's a Guru3D review of the 6800K where they pair both with a GTX 580--when the CPU is the limiting factor the 4770K is about twice as powerful), so in order for that DirectX graph to make sense we would need to expect a near 75% gain in per thread performance from Richland, which is not possible.

 

[kn00tcn]

In CPU limited situations under DirectX the 4770K performs about twice as well as the highest end Richland CPU (there's a Guru3D review of the 6800K where they pair both with a GTX 580--when the CPU is the limiting factor the 4770K is about twice as powerful), so in order for that DirectX graph to make sense we would need to expect a near 75% gain in per thread performance from Richland, which is not possible.

what about cpu only tasks like cinebench or known single threaded ones like lame encoder

i guess i'll look around, or we'll have more info in the coming weeks as more tests happen

 

[HD64G]

In CPU limited situations under DirectX the 4770K performs about twice as well as the highest end Richland CPU (there's a Guru3D review of the 6800K where they pair both with a GTX 580--when the CPU is the limiting factor the 4770K is about twice as powerful), so in order for that DirectX graph to make sense we would need to expect a near 75% gain in per thread performance from Richland, which is not possible.

What are you talking about? Kaveri's iGPU vs 4770's iGPU isn't in your calculations? Why? CPU's bottleneck is almost diminished in Mantle, so iGPUs are the ones that work most. So, Kaveri's iGPU matches 4770's in this demo.

 

[The Von Matrices]

What are you talking about? Kaveri's iGPU vs 4770's iGPU isn't in your calculations? Why? CPU's bottleneck is almost diminished in Mantle, so iGPUs are the ones that work most. So, Kaveri's iGPU matches 4770's in this demo.

That benchmark is not comparing iGPUs.

Even the "old" Trinity based A10-5800K beats the 4770K at iGPU performance in DirectX. Also, you can't run Mantle code on a Intel iGPU, unless the world has begun to end and Intel adopted Mantle.

 

[xenocide]

what about cpu only tasks like cinebench or known single threaded ones like lame encoder

i guess i'll look around, or we'll have more info in the coming weeks as more tests happen

Here's the review I was looking at: http://www.guru3d.com/articles_pages/amd_a10_6800k_review_apu,1.html

There are very few situations where the CPU alone performs anywhere near what an i7-4770K can do.

That benchmark is not comparing iGPUs.

Even the "old" Trinity based A10-5800K beats the 4770K at iGPU performance in DirectX. Also, you can't run Mantle on a Intel iGPU, unless the world has begun to end and Intel adopted Mantle.

Exactly. Which means they are measuring exclusively CPU performance, and as the link above shows, there's nearly no way it would be that close without some fudging.

 

[HD64G]

That benchmark is not comparing iGPUs.

Even the "old" Trinity based A10-5800K beats the 4770K at iGPU performance in DirectX. Also, you can't run Mantle code on a Intel iGPU, unless the world has begun to end and Intel adopted Mantle.

And how do you explain 4770 increase from 12-13 to 32-33? And the graph label saying no API overhead?

What isn't normal here is that Intel CPU-APU owners trying to fight the proofs of something they should wish. Mantle's ability to diminish the CPU load in games. Hey guys, don't you wish your 8-threaded cpus would do more things at the same time except for losing cycles because of directX's inefficiency?

 

[The Von Matrices]

And how do you explain 4770 increase from 12-13 to 32-33? And the graph label saying no API overhead?

You use the same discrete GPU in both systems?

 

[HD64G]

You use the same discrete GPU in both systems?

Correct. But then AMD would crossfire if some GPUs were used. Do we have the system specs to clarify this?

 

[The Von Matrices]

What is normal here is that Intel CPU-APU owners trying to fight the proofs of something they should wish. Mantle's ability to diminish the CPU load in games. Hey guys, don't you wish your 8-threaded cpus would do more things at the same time except for losing cycles because of directX's inefficiency?

This would be nice, but it's not happening in this benchmark. If the benchmark could use 8 threads, then we would expect ~80% more performance from the Intel CPU since it has double the number of cores/threads. This is why people like I are skeptical about these results, because a threaded workload would still heavily favor the Intel CPU. It's almost as if this benchmark is constrained to 2-4 threads and is using a weak discrete GPU in order to shine the best light on the APU as possible. Note, I'm not doubting that Mantle can improve performance, just the benchmark's claim that Mantle makes CPUs of different performance tiers in other applications perform equally in games.

Correct. But then AMD would crossfire if some GPUs were used. Do we have the system specs to clarify this?

I doubt we'll get that information until after the NDA of January 14. However, I also doubt Crossfire with the iGPU was used; after all, wouldn't the AMD system be on top if it had two GPUs instead of one? That wouldn't be a very good PR slide if the advertised system with double the GPUs still performed worse than the competition.

 

[xenocide]

I know Mantle has the potential to cause great performance gains, but their numbers don't add up, and they are throwing out astronomical values left and right with nothing but slides to back it up.

 

[the54thvoid]

If Mantle uses my cpu's 12 threads of awesomeness I'm going to buy a matchbox filled with resistors to do my gfx work.

 

[RoutedScripter]

I doubt we'll get that information until after the NDA of January 14. However, I also doubt Crossfire with the iGPU was used; after all, wouldn't the AMD system be on top if it had two GPUs instead of one? That wouldn't be a very good PR slide if the advertised system with double the GPUs still performed worse than the competition.

That's because i7-4770 is in 350$ price range while A8-7600 is below 120$. yo yo yo

http://www.newegg.com/Product/Product.aspx?Item=N82E16819116901


http://i.imgur.com/8axFUzi.jpg

 

[TheMailMan78]

Pixie dust and black magic.