Media Encoding and OpenCL

Media Encoding

Cyberlink MediaExpresso 6.5 and Arcsoft MediaConverter 7.5 are two softwares that are currently supporting the Intel Quick Sync and AMD VCE. When comes to video transcoding, nothing can touch Intel’s QuickSync at the moment. The A10-5800K is just as fast as the Core i5 2500K without any hardware acceleration and is faster than the Llano. With VCE enabled, we shed 3 seconds off compared to the fast mode and 27 seconds off compare to the quality mode.

HandBrake: OpenCL Performance

HandBrake is an open source media converter that supports OpenCL. The software is available online free to download but the current version 0.9.8 does not yet support the OpenCL and we have gottena  beta version that allows us to test the performance with and without OpenCL. We did not have much time to run the test through a list of CPUs so we only include a small data set.

Clearly, A10-5800K is slower than the Sandy Bridge CPU when transcoding video through the x86. With OpenCL enabled, we got 54% gain in performance. We will have to do more test and compare the performance with the GPU acceleration on Sandy Bridge and OpenCL performance of the Ivy Bridge. So far, we can see that if the software is capable to tapping into the GPU power via OpenCL, Trinity can gain significant gain where it goes from being underperforming be a very competitive at its price range. 

Ultimately, this all lies on the software support. Currently, AMD lists about 10 applications that supports OpenCL that includes some of commonly used software titles: WinZip 16.5, GIMP 2.8, Adobe Flash 11.2, Adobe Photoshop CS6, ArcSoft MediaConverter 7.5, and CyberLink MediaExpresso 6.5, Handbrake. Obviously, software programmers would need to take their time and implement such support and we hope more software will adopt OpenCL and similar technology to fully take advantage of modern processing power. 

Compute Power: OpenCL and OpenGL

Let’s take a look the theoretical GPU compute power of the HD 4000. We ran Sandra’s GP Computing  benchmarks. Keep in mind that these numbers are theoretical performance.

GP Processing

OpenCl GP Processor

OpenCl GP	Native Float Shader (Gbpix/s)	Native Double Shader (Mpix/s)
Core i5 2500K (HD 3000)	0.10388	43.3
Core i7 3770K (HD 4000)	0.15791	72.75
A10-5800K (GPU only)	0.6425	42.8
A10-5800K (GPU+CPU)	0.65528	27.65
A8-5600K (GPU only)	0.4102	57.3
A8-5600K  (GPU+CPU)	0.42335	48

The theoretical OpenCL compute power on the A10-5800K is about six times greater than the Core i7 3770K in the native float shader (single precision) benchmark. However, the A10-5800K falls slightly behind the Core i5 2500K in the Native double shader (double precision) benchmark. 

Sandra also provides the ability to measure the CPU and GPU combination performance and with the 

Compute Shader	Native Float Shader (Gbpix/s)	Native Double Shader (Mpix/s)
Core i7 2500K (HD 3000)	0.05881	5
Core i7 3770K (HD 4000)	0.27276	57.2
A10-5800K (GPU)	0.57082	24.43
A8-5600K (GPU)	0.382	16

Looking at the Computer Shader, we see the A10-5800 is twice as fast as the Core i7 3770K in the single precision float point calculation but it is only half as fast in the double precision calculation. The A10-5800K is, however, much faster than the Core i7 2500K.

GP Cryptography

Compute Shader	Cryptography Bandwidth (GB/s)	Encryption/Decryption Bandwidth (GB/s)	Hashing Bandwidth (GB/s)
Core i7 2500K (HD 3000)	0.363	0.129	1.019
Core i7 3770K (HD 4000)	1	0.902	1.15
A10-5800K (GPU)	2.2	1.15	4.23
A10-5800K (GPU+CPU)	2.37	1.24	4.5
A8-5600K (GPU)	1.67	0.76	3.77
A8-5600K (CPU+GPU)	1.8	0.971	3.64

In the Cryptography benchmark, the Trinity trumps the fastest Ivy Bridge 

GP Bandwidth

Compute Shader	Internal (GB/s)	Transfer (GB/s)
Core i5 2500K (HD 3000)	8	3.29
Core i7 3770K (HD 4000)	14.77	6.7
A10-5800K (GPU)	11.52	6.8
A8-5600K (GPU)	11.47	6.73

Looking at the compute shader bandwidth, we an see the A10-5800K is faster than the Core i5 2500K by 50% in the internal and 100% in the transfer. However, comparing it against the Core i7 3770K, it falls behind in the internal but is able to keep up with the transfer.