New patent presented by AMD for its graphics cards, would partly solve the excess consumption that they currently present.
There has been a lot of talk and talk about the consumption problems of AMD graphics cards. Polaris is quite inefficient but it was made up for by acceptable performance and a reasonable price. Vega has been a complete disaster, with serious problems of consumption, temperature and a very high price. The problem is in the Graphics Core Next (GCN) architecture, which has design deficiencies that cause heavy consumption. To correct the problem they would have developed a system that improves it and which they have patented.
AMD's consumption problem
They have filed the patent: "High bandwidth data stream processor and low power vector log file." This patent aims to correct the high consumption in future graphics cards. Although it was published on December 13, 2018 by the United States Patent Office, now it is when it has been made public. This refers to a parallel processing unit similar to that used in current AMD, Intel, NVIDIA, or ARM architectures.
AMD seeks with this patent to solve the performance problem when resources such as the cache are shared between the Stream Processors of the GPU. These cores could execute different types of operations in parallel. In the event of incorrect access to information, moving another record involves additional consumption by the GPU.
Optimize the operation and parallel execution, allowing to share resources efficiently or reduce those that are shared. This offers to minimize access to memory and movements between registers, especially to reduce consumption. The energy required for these movements is studied extensively and is measured in nanojoules per movement. The current execution path in which the order of actions is established in parallel is always the same and therefore, it is inefficient.
Solution posed by AMD
The patent seeks to change the vector log file system by increasing bandwidth. This allows for greater speed in record movements. It also offers a Vector Arithmetic Logic Unit (VALU) with shading changes in the form of new instructions. They are also implemented in adding more memory and other system input / output changes. This is presented in detail in a diagram of the instruction execution process.
Possibly this improvement was already implemented in Navi, although there is no record of this. What we denote as positive is that if it becomes a more efficient process, it will be positive. AMD had to attack the problem urgently and it seems that they have done so and have found a solution.
Source: HH
