AMD vs Intel Performance With OpenCL Realtime Face Animation

[fusion238]

Last week, AMD Ventures company Mixamo announced software utilizing OpenCL based realtime capture and face animation. A chart comparing AMD and Intel's Haswell-based i5-4670 ($215 CPU, 84 watts TDP) shows the 35w mobile AMD APUs outperforming the desktop Intel CPU.

This foreshadows what kind of performance AMD's upcoming Kaveri APU with GCN engines will deliver to the market in the near future.

http://community.amd.com/community/...us-plug-in-for-the-unity-development-platform

 

[piesquared]

The heterogeneous approach is clearly the path going forward and almost the entire industry is saying the same thing. It makes complete sense that software should target IP blocks that processes the particular data in the most efficient and highest performing way. And this example isn't even using hUMA yet, with already a 2X advantage over competing products.

 

[mikk]

Surely optimized for AMD, so no wonder. A8-4500M is much slower both in gaming and OpenCL.

 

[TerryMathews]

Surely optimized for AMD, so no wonder. A8-4500M is much slower both in gaming and OpenCL.

You make it sound as though Intel would never optimize their software to the detriment of AMD.

 

[Phynaz]

Could you post the * please?

 

[TerryMathews]

Could you post the * please?

* Tests performed by AMD. PC manufacturers may vary their configuration yielding different results. Benchmark and test files provided by Mixamo. Test project measures the average speed at which the system can apply the Mixamo Face Plus facial gesture recognition and animation algorithms using a 480p test video containing a full range of facial emotions and gestures that is loaded into system memory to simulate a live video feed from a standard webcam in an uncontrolled environment. An HP ProBook 6475b notebook PC with AMD A10-4600M APU with AMD Radeon™ HD 7660D Graphics, 8GB DDR3-1600 RAM, video driver 9.12.0.0, 12/19/2012, Windows 7 Professional 64-bit (6.1, Build 7601) Service Pack 1 achieved 42 frames per second with GPU acceleration on, 3 frames per second with GPU acceleration off. TRN-201

With an AMD reference desktop PC configured with an AMD A10-6800K APU with AMD Radeon™ HD 8670D Graphics, 8GB DDR3-2133, video driver 13.101.0.0, 6/4/2013, Windows 7 Professional 64-bit (6.1, Build 7601) Service Pack 1 achieved 49.5 frames per second with GPU acceleration on. An ASUS All Series desktop PC configured with an Intel Core i5-4670 with Intel HD 4600 graphics, 8GB DDR3-1600, video driver 9.18.10.3165, 5/7/2013, Windows 7 Professional 64-bit (6.1, Build 7601) Service Pack 1 achieved 22 frames per second with GPU acceleration on. RID-30

An HP Pavilion dv6 notebook PC with AMD A8-4500M with AMD Radeon™ HD 7640G Graphics, 6G DDR3-1600, video driver 12.104.0.0, 3/28/2013, Windows 7 Home Premium 64-bit (6.1, Build 7601) Service Pack 1 achieved 30.9 frames per second with GPU acceleration on. TRN-202

 

[AnandThenMan]

Surely optimized for AMD, so no wonder.

Exactly. Just like half the benches out there are Intel optimized, or even have code supplied by Intel. I keep saying, AMD has to make their mark by ramping up the software side of things. The best hardware in the world does no good if the software is not there to take advantage.

 

[Enigmoid]

Personally I'm a little leery of this release. AMD has fudged the numbers on their marketing slides before (didn't they not release a slide before where the i5 was running OpenCL but on the CPU cores?). 13x is a massive increase and I can't seem to find a place where I have ever seen numbers than high. That's high even for quicksync and that has custom silicon.

I'm not even sure why they used the test they did.

According to Mixamo, for cameras capable of higher frame rates and larger resolutions, the increased amount of data needing to be processed is offset by the higher quality video such that the overall processing requirements compared to a poor quality 480p video are lower. Hence, using a poor quality, SD video to simulate something closer to a worse case scenarios was used, which enabled consistent testing to be applied across all parts tested.

Why on earth would a dev ever use such a poor quality cam?

Confusingly it looks like you could run high bitrate 720p ot 1080p video and get better numbers.

 

[Phynaz]

* Tests performed by AMD.

Thanks, all I needed to know

 

[mikk]

You make it sound as though Intel would never optimize their software to the detriment of AMD.

A8-4500M is a slow APU, in basically all OpenCL applications way slower than IVB or HSW. There is something going on in this bench. But this is expected from a marketing slide.

 

[TerryMathews]

A8-4500M is a slow APU, in basically all OpenCL applications way slower than IVB or HSW. There is something going on in this bench. But this is expected from a marketing slide.

This is simple, I don't know why you can't see it; it's right on the slide. They forced rendering through the GPU.

A8-4500M might be slow, but the Radeon in it is faster than the IGP in the i5. Whether or not that is a fair test (we both know it is the best case for AMD and the worst case for Intel), it is "apples to apples".

 

[fixbsod]

8 gb 1600 for intel and 8 gb 2133 for amd...

 

[monstercameron]

8 gb 1600 for intel and 8 gb 2133 for amd...

no scandal here just max supported memory
http://ark.intel.com/products/75048/

 

[fixbsod]

I'd still like to see a more apples to apples

no scandal here just max supported memory
http://ark.intel.com/products/75048/

 

[monstercameron]

I'd still like to see a more apples to apples

different products deal with memory differently...so you cant get direct comparisons like that...er besides this is apples to apples, stock amd v stock intel

 

[BallaTheFeared]

Looks good, I have no interest in using AMD's software or using software such as this though.

I'm sure someone will appreciate it.

 

[SiliconWars]

Thanks, all I needed to know

You could always ask Anand to test it as well.

 

[Idontcare]

8 gb 1600 for intel and 8 gb 2133 for amd...

no scandal here just max supported memory
http://ark.intel.com/products/75048/

Quite a gap in price for the memory though, which matters when making statements like "NOT possible with comparably priced Intel products".

 

[Enigmoid]

This is simple, I don't know why you can't see it; it's right on the slide. They forced rendering through the GPU.

A8-4500M might be slow, but the Radeon in it is faster than the IGP in the i5. Whether or not that is a fair test (we both know it is the best case for AMD and the worst case for Intel), it is "apples to apples".

Intel is very competetive in igp compute with the HD 4600 generally competing with or beating the a10-6800k in OpenCL performance. Considering the HD4600 is better than or equal to the a10-4600m in games its fair to say that the a8-4500m should not be leading and by that margin.

 

[TerryMathews]

Intel is very competetive in igp compute with the HD 4600 generally competing with or beating the a10-6800k in OpenCL performance. Considering the HD4600 is better than or equal to the a10-4600m in games its fair to say that the a8-4500m should not be leading and by that margin.

You have no basis to make that statement. You have no idea what instructions this program is using and the performance delta of those instructions between Intel and AMD.

 

[Khato]

Intel is very competetive in igp compute with the HD 4600 generally competing with or beating the a10-6800k in OpenCL performance. Considering the HD4600 is better than or equal to the a10-4600m in games its fair to say that the a8-4500m should not be leading and by that margin.

As another already stated, it's simply due to differences in optimization. OpenCL GPU implementations have a tendency to have 'glass jaws' that need to be avoided in order to obtain peak performance. This appears to be hitting such an issue on the Intel HD 4600. All the more reason for AMD to make a marketing slide out of it... just like Apple used to find obscure benchmarks to trumpet how much faster their G4 powermacs were compared to PCs when the opposite was true in most cases.

 

[Enigmoid]

You have no basis to make that statement. You have no idea what instructions this program is using and the performance delta of those instructions between Intel and AMD.

As another already stated, it's simply due to differences in optimization. OpenCL GPU implementations have a tendency to have 'glass jaws' that need to be avoided in order to obtain peak performance. This appears to be hitting such an issue on the Intel HD 4600. All the more reason for AMD to make a marketing slide out of it... just like Apple used to find obscure benchmarks to trumpet how much faster their G4 powermacs were compared to PCs when the opposite was true in most cases.

Its unusual in that no other OpenCL benchmark show such a difference.

http://www.tomshardware.com/reviews/core-i7-4770k-haswell-review,3521-3.html

HD 4600 beats richland every single time.

http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/17

On the openCL tests the two are generally equal.

You can look in a lot of benchmarks but generally the a10-6800k never beats the hd 4600 by more than a factor of 2.

This is a promo slide. Does anyone ever take those at face value?

 

[TerryMathews]

Its unusual in that no other OpenCL benchmark show such a difference.

http://www.tomshardware.com/reviews/core-i7-4770k-haswell-review,3521-3.html

HD 4600 beats richland every single time.

http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/17

On the openCL tests the two are generally equal.

You can look in a lot of benchmarks but generally the a10-6800k never beats the hd 4600 by more than a factor of 2.

This is a promo slide. Does anyone ever take those at face value?

It's too specific to be false. AMD laid out configurations for both test rigs down to the OS version. It would be simple for Intel to duplicate, and Intel would have a claim of action if the ad were demonstrably false.

Look, I agree with you. This is an engineered benchmark designed to make AMD look better than Intel. I don't understand why you find it so difficult to believe that a subsidiary of AMD could create a piece of software that performs better on AMD's APUs than Intel's Core chips?

 

[AtenRa]

Intel is very competetive in igp compute with the HD 4600 generally competing with or beating the a10-6800k in OpenCL performance. Considering the HD4600 is better than or equal to the a10-4600m in games its fair to say that the a8-4500m should not be leading and by that margin.

What games have to do with this application ?? Games use the CPU as well, this software is using the iGPU only. Also it is heavily optimized for the iGPU (VLIW4) in Trinity/Ritchland APUs. Like Cinebench is optimized for Intel CPUs.

Also, there are more applications that AMD APUs have way higher performance than Intel APUs like Musemage, Photozoom, photoshop(some filters) etc. As for the CLbenchmark 1.1.3, it uses both the CPU and GPU, it is why the Core i7 4770K(87W) is a little faster than A10-4600M(35W). Same happens with Sony Vega and more.

Edit: Just to add,
A8-4500M has 2/3 the Radeon cores (256) than A10-4600M (384). That is also almost the difference of performance in the slide.

 

[ShintaiDK]

Nice PR propaganda thread. A more unbiased source couldnt be found. And funny how it doesnt reflect regular reality and products

Its amazing how many times people burned their fingers, and still havent learned.