DirectX 11 Compute Shader: Three times faster than DX10.1 due to Local Data Share
AMD's Developer Relations boss Richard Huddy explains the mode of operation of Ambient Occlusion - the DirectX 11 Compute Shader is said to provide up to three times the performance possible with DirectX 10.1.
Using a Compute Shader path in DirectX 11 could offer up to three times the performance possible with DirectX 10.1, mentions Richard Huddy, AMD Developer Relations, in an interview with Pc Games Hardware. And even in DirectX 10.1 modern Radeon cards with their Fetch4 feature are able surpass pure DirectX 10 devices like the current Geforces.
But at the same time Huddy explains that Compute Shader is uncharted territory for developers and thus it is hard to integrate. Getting the optimum out of it is even harder currently.
In matters of possible performance benefit Huddy calls the Local Data Share (the memory shared between different Threads), which is required by DirectX 11, a Killer-Feature. Especially in Screen Space based Ambient Occlusion it is possible to safe huge amounts of data traffic resulting in less workload for the bandwidth and texture units which are used to read the data.
PC Games Hardware has already reported about AMD's DirectX 11 graphics cards of the Evergreen family, which are said to be released in 2009. We have integrated the first video with Richard Huddy's interesting remarks about the expected performance into this article again.
As it has already been the case at the Computex AMD concealed the real performance of the graphics card - accordingly the fps values visible on the display are not representative for the final products. The Evergreen card with a dual slot cooling solution is, according to Huddy, neither an entry level nor a high-end product - it is supposed to be part of the product array above 100 USD.
Grafikkarten-Rangliste 2016: 32 Radeon- und Geforce-GPUs im Benchmarkvergleich [Oktober]