AMD on asynchronous shaders in the Radeon series and the benefits of DirectX 12

The introduction of Mantle accelerated Microsoft and Khronos Group to develop new interfaces that allow developers to work closer to the hardware, something that, among other things, provides better opportunities to use more processor cores (threads) and lower overhead. Now AMD is revealing other benefits of the upcoming tools, not least in combination with Radeon graphics cards.

The ability to work closer to the hardware brings other benefits than using more processor cores. In a conference call, AMD says that with DirectX 11, things like graphics computing, general calculations, and memory management happen in series, something that involves poor hardware utilization and lower frame rates.

The fact that a lot is managed serially is attributed to the fact that graphics processors as a result of the currently common interfaces are a “black box” for game developers. With DirectX 12, there are greater opportunities to parallel the work, something AMD claims will contribute to a shorter time being needed for each rendered frame.

However, AMD believes that parallelization poses other challenges at the hardware level, where some graphics processors can only process one stream of commands at a time. This means additional complexity in scheduling and making full use of the hardware, something that can be solved to some extent by pausing a power while a task that has a higher priority takes precedence.

Here, AMD is said to have an advantage with what is called Asynchronous Compute Engines (ACE), which was introduced with the architecture Graphics Core Next (GCN) at the end of 2011. It is about dedicated resources in the graphics processor that make it possible to handle several simultaneous command streams. This should be extra important or even necessary for lower latencies with virtual reality, according to AMD.

The number of ACEs varies between different graphics processors, with Hawaii in the Radeon R9 290X and R9 290 having the largest number with 8 pieces. Each unit can also handle up to 8 simultaneous queues, or a total of 64. This is a feature AMD believes can provide increased performance with “advanced” graphics effects and in addition, competing graphics solutions are said to lack anything similar.

At present, there is only limited functionality for asynchronous calculations with DirectX 11 and AMD’s application interface Liquid VR for virtual reality. Full support and not least the fact that the company’s ACE units can be used in games is delayed until the launches of DirectX 12 and Vulkan.