OverclockingThe AMD A-Series processors do not feature an unlocked multiplier, even though some BIOS screens might suggest it. When you change the multiplier upward on an A-Series CPU, the BIOS, Windows and even tools like CPU-Z will show the changed multiplier, but the change will not actually have any effect on the actual clock frequency of the processor. That's why we recommend changing the bus clock only.
On the Llano platform, CPU and GPU clock are all based on the PCI-Express frequency, which means that certain clock choices might result in a PCI-E or SATA clock that is far enough out of spec to make certain components stop working. At certain points, depending on motherboard hardware and BIOS design, additional dividers become active, which means that you should try higher clocks, even if the current one is unstable.
We tested the FPS improvement from changing bus clock only, using Unigine Heaven. As you can see the performance increase is almost linear, around 0.13 FPS extra per 5 MHz bus speed. This is a radical difference to what you would see on a "normal" CPU where the CPU frequency is completely independent from the GPU clock. In that case only a minimal increase in FPS would become apparent.
Once we went beyond 120 MHz, the analog VGA output of the board stopped working and we had to switch to DVI to continue overclocking. Beyond 130 MHz, our CPU's overclocking capabilities were maxxed out when using the x29 multiplier, so we had to reduce it to x27 to reach our maximum of 135 MHz.
On one hand these interlocked frequencies make it easy for novices to increase total system performance because they have to tweak only one parameter. On the other hand the dependency can result in multiple components being the limiting factor, untangling which requires a deep understanding of clock interactions - or trial and error.
UMA Memory Buffer SizeThe integrated graphics of all AMD A-Series uses the system's main memory as graphics memory. There are two ways how (system) memory can be allocated as graphics memory: UMA (Unified Memory Architecture) and AMD HyperMemory.
The allocation size for UMA is set in the BIOS and dedicates that amount of memory to the GPU, away from the CPU and Windows will have no access to this memory slice at all. Once UMA memory runs out, AMD's driver has a feature called HyperMemory which lets the driver allocate additional memory from within Windows, like any other application would do, to be used as graphics memory.
UMA has the advantage that the graphics hardware has access to a single, dedicated, unfragmented memory range that it can work its magic in without worrying about the rest of the system. Here, the disadvantage is that the memory is assigned to the graphics chip, even when the graphics hardware does not need that much memory, for example when only the Windows Desktop or 2D applications are active.
HyperMemory solves the problem of unused memory by dynamically allocating as much memory as the GPU needs, when it needs. As a result more (system) memory is available for normal applications to use, but performance of the graphics memory is reduced. Since the driver has to manage this memory allocation, which is in turn passed on to the Windows Kernel, several layers of processing are added which results in a significant performance impact as we see in the following results.
Again, we tested Unigine Heaven with all UMA memory size options available in BIOS. As you can see there is a well visible performance impact from smaller sizes. However, even with 32 MB, graphics performance is reduced by only 10%. So in scenarios where the system is spending most of the time at the desktop or doing 2D work, it might be beneficial to use a smaller UMA size, which frees up memory for Windows that would otherwise be allocated to the GPU in a fixed manner and end up being unused.
For a machine that favours gaming or that has lots of memory installed anyway, a UMA size of 512 MB provides maximum performance, yet will reduce available memory in Windows by 512 MB.
Memory TimingsAs mentioned before, AMD's integrated Radeon Graphics in the A8-3850 uses the motherboard's main memory as graphics memory which means that graphics performance is directly dependent on this memory's performance characteristics.
Our testing, with Unigine again, shows a very distinct performance effect of increased memory clock frequency. It looks like the sweet spot is at 1600 MHz, at which point memory is still extremely affordable. If you need to maximize performance, opting for the 1866 MHz memory variant offers around 5% extra 3D rendering performance. Certainly interesting, but not worth it when taking into account price/performance, as 1866 MHz modules are far more expensive than 1600 MHz memory.
Memory timings play only a minor role, so it is better to maximize memory clock frequency first, and then look at timings, not the other way round. Another important result here is that running the memory at the faster CL6 timing does actually reduce performance when compared to CL7 or CL8.