More and more details about the new architecture of AMD Navi GPUs are gradually being revealed. As you know, she will the long-used Graphics Core Next (GCN) architecture, but at the same time, according to the latest data, it will receive very tangible changes. In particular, the new architecture will correct one serious flaw inherent in previous versions of GCN.
The web was conditional diagram of the Navi 10 GPU. Judging by it, the computing power of the GPU will be divided into eight shader units (Shader engine), each of which will have five computing units (Compute units, CU). Recall that each CU in the GCN architecture has 64 stream processors. That is, in total, Navi 10 will have 2560 stream processors. Quite good for a mid-range chip.
Previous versions of the GCN architecture assume that the compute units are divided into four shader units. Such an arrangement was, for example, in the Hawaii Pro chip with the same 2560 stream processors, and it remained the same in the Polaris and Vega GPUs. And that's why AMD GPUs couldn't offer more than 64 raster operations units (ROPs).
Therefore, the decision to divide the Navi GPUs into eight shader units gives hope for a doubling of the number of raster operations units, that is, there will be 128 of them. This, of course, should have a positive effect on the performance of the GPU, especially in games.
Source: 3dnews.ru