Back in September, AMD launched their latest 7th Generation A-Series Bristol Ridge APUs, as well as the new AM4 platform. The official launch only announced the availability of the BR APUs for OEM systems, the DIY retail launch is planned for a later to be announced date. This was a bit of a letdown since the Bristol Ridge pre-announcement took place in April and we saw no reason why system builder couldn't get their hands on the new platform this year. Naturally, we are eager to test out the newest APU offerings and can't really wait any longer for the PIB (Product in Box) release, so we decided to evaluate a preconfigured OEM build - HP Pavilion 510-p127c with the A12-9800 APU.
The AM4 socket will support both the upcoming Zen based CPUs and Excavator based BR APUs, so this is a great milestone for many upcoming AMD products. The most important update for the APU platform is the support for DDR4 memory kits - up to 2400 MHz. As AMD APUs are generally bandwidth limited in graphics based workloads, this could result in huge performance gains.
As a reference, we will compare the HP OEM system to our InWin SFF A10-7860K and i3-6100 rigs.
- As our main goal is to investigative the capabilities of the APU, we have removed the dGPU from the OEM system.
- Unfortunately, the BIOS shipped with the Pavilion is really basic, we could not find any way to increase the dedicated memory size of the graphics processor, by default it is 512 MB.
- We always aim to use the latest drivers possible from every vendor, this time around AMD Crimson 16.40.2311 & Intel 126.96.36.19942 were used in the tests.
In this session we will have a look at traditional CPU benchmarks.
Hyper Pi is tool that can calculate the value of pi to a desired number of decimal places. It can spawn multiple instances of Super Pi processes to stress test a PC to determine its overall performance and verify its stability. We calculate the value of pi to 2M digits using one thread only.
7-Zip is an open-source file archiver with high compression ratio. We measure how fast it can compress the Thai Statue 3D file from The Stanford 3D Scanning Repository in multi-threaded mode.
CINEBENCH is a popular CPU benchmark from the creators of Cinema 4D suite. The tool renders a photo-realistic 3D scene, it supports both single-threaded and multi-threaded stress tests.
Blender is a professional free and open source 3D content creation suite. Blender Institute, the studio that organizes the Blender Foundation goals, has prepared six Blender scenes for benchmarking the Cycles render engine - we picked bmw27, classroom and Pabellon Barcelona scenes. Cycles is Blender’s ray-trace based production render engine that support CPU and GPU rendering (CUDA and OpenCL) as well. Unfortunately, we didn't succeed in setting up Cycles for GPU rendering on integrated graphics (Intel is not supported and AMD didn't render anything) and therefore we had to resort to traditional CPU rendering. Blender is cross-platform so it would be interesting to investigate how rendering times change on different OSes - we run Blender without GUI on Windows 10. To keep render times in check, we lowered the sample numbers from their default values and set the resolution to 1080p with 50% scale for each scene.
The gathered data shows around 10% difference in favour of the new platform, with the biggest gains in HyperPi. Blender's ray-tracer really shines on the i3-6100 with an average advantage of a staggering 80%, while in Cinebench the gap between BR and Skylake is 27%.
Two synthetic GPU benchmarks are deployed to measure the performance of the integrated graphics solutions.
LuxMark is the de facto OpenCL based ray tracing benchmark. Ray tracing was on of the first areas where GPGPU based acceleration techniques became really popular and effective. LuxMark is an easy to use, open source cross platform benchmark, based on LuxRender. We use all the available test scenes to collect performance numbers. Although OpenCL can target all available devices in the system (or even mix them) - we only select the iGP as the accelerator for ray tracing computations.
ShaderToy is a popular platform for creating intricate procedural worlds and "demo scene" effects. The scenes are coded in complex fragment shaders written in GLSL and often adopt ray-marching techniques to generate the final image - with long loops and lots of ALU operations. We picked 3 stunning looking scenes created by Inigo Quilez - a well known ShaderToy artist. We run the shaders for 30 seconds and report the average frame rate for the scenes.
The recently released graphics driver 188.8.131.5242 from Intel, which is based on a whole new code-branch, brought huge gains (from 1.25x to nearly 2.0x) in these benchmarks for the i3 - older driver was 184.108.40.20639. This boost means that the i3 leapfrogged AMD APUs in LuxMark scenes and closed the gap considerably in ShaderToy. BR's 30% GPU clock speed advantage over the Kaveri system really helps in ShaderToy scenes.
All in all, Bristol Ridge has some nice improvements over Kaveri, but this won't change things that much regarding the performance delta that AMD holds in CPU based workloads compared to the i3. Bristol Ridge is all about power efficiency compared to the previous offerings, with the additional connectivity options (USB 3.1, NVMe, SATA 6.0) and DDR4 memory support AMD could score more design wins in the all important OEM business.
Sadly, game benchmarks haven't shown any performance difference between the two generations of AMD APUs. We need some more time and have to dig deeper to see how different factors influence frame times on Bristol Ridge, but being a mobile focused platform, our gut feeling is without additional DDR4 OC AMP profiles it will be hard to find any gains in gaming scenarios. Moving forward, we would like to do in-depth power usage benchmarks and build our own BR rig as well.
Coming January, both AMD and Intel plan to release new CPUs to the market, ZEN and Kaby Lake will most likely add more options and opportunities ... interesting times ahead!
Correction: Messed up the CPU freq for the BR APU on the build slide and left to the charts in the video: it's 3.8 GHz, not 4 GHz. Sorry for the typo!