Wednesday, June 15, 2011

AMD leaves the basic architecture of VLIW

It ended the first day of 'AMD Fusion Developer Summit 2011. THE most important conference on the stage has seen Michael Mantor - AMD Senior Fellow Architect - Houston and Mike - AMD Fellow Architect - speak of the future architecture of AMD GPUs. What AMD has not shown the full details of a particular generation of graphics chips, but a preview of the architectural features behind these solutions.

At the moment it is unclear whether the adoption of this new type of architecture will be realized as early as next year with the Radeon HD 7000 (Southern Islands?) Or if it is a project only on paper with a longer timeframe. If the GPU Cayman we have seen AMD try to change and improve static type of architecture carried out by five generations (R600 ~ EG / NI), with the GPU "next-gen" (used to indicate quest'abbreviazione architecture which do not yet know the code name), we see a clean break with the past.

In an effort to simplify programming and move the main load on the hardware side, AMD has decided to abandon the basic VLIW architecture in favor of a "Scalar + Vector" in some way closer to the vision of the second GPU nVidia. The slide below shows the different organization of the shaders on GPUs and GPU Cayman "next-gen": Shoot the SIMD-Engines in favor of what AMD calls Compute Units (CU), similar to Streaming Multiprocessor (SM) from Nvidia.

If the SIMD single-engine Cayman VLIW is composed of 16 x 4, the new CU un'ALU superscalar SIMD 4 independent units. In the development of new shared core AMD has redesigned all the parts of a graphics chip, including the front-end processor and the command structure of the cache, all changes that serve to increase the performance and efficiency in multi-tasking.

To this are introduced Asynchronous Compute Engines (ACE), entities that have the task of picking operations without trial, the graphic part. Unlike nVidia, AMD continues to perform geometric calculations, and HOS tessellation separately but now the entire management eventually find highly parallelized.

It 'important to remember that despite the development of AMD's new architecture will not give support to the implementation VLIW4 (GPU and driver), at least not in the immediate future, since the APU Trinity - always shown all'AFDS and expectation in the market during 2012 - will benefit from an offset iGPU Cayman.

No comments:

Post a Comment