AMD Announces Vega Architecture Details
AMD has announced details about the Vega architecture during CES, which will be found in its new video cards. The emphasis is on faster HBM2 memory that is also used more efficiently than with previous architectures.
AMD has not released anything about clock speeds, compute units or consumption figures, but the manufacturer has released details about the architecture of its new GPU codenamed Vega. The emphasis is on memory, because so far the growth in computing power has been relatively greater than the growth in memory capacity, according to AMD.
The new architecture therefore uses hbm2, whose maximum capacity is 32GB, while the bandwidth doubles compared to the first generation hbm. As a result, the memory bandwidth can reach up to terabytes per second. Whether AMD will actually provide a video card of 32GB hbm2 is a matter of waiting, because no statement has yet been made about such matters.
In addition to fast memory, efficient use of memory is also important according to AMD. To keep memory usage in line, AMD has developed the ‘high-bandwidth cache controller’, which sits on the GPU and manages the virtual address space. The hbcc can address a maximum of 512TB, spread over video memory, working memory and storage such as an SSD or hard disk. The controller ensures that data that the GPU needs is stored in the fast memory, while data that does not have to be immediately available is parked elsewhere. AMD’s own tests, for example, would show that in Fallout 4 and The Witcher 3 roughly only half of the memory that is allocated is actually used and so memory can be freed up to use for other things.
AMD has also tackled the GPU itself. There is support for primitive shaders, which can take over the work of a vertex and geometry shader and run faster. This can be done at the driver level, where the driver determines when that is possible, or that is possible when support for this is built into a game. That would mean doubling the speed per clock tick.
Furthermore, AMD has improved the load balancing and the compute units have been improved. This allows them to operate at higher frequencies and also carry out more instructions per clock tick. In addition, memory compression has been improved and render backends have now become clients of the l2 cache, providing more speed for applications using deferred shading.
It remains to be seen what the Vega graphics cards will look like and what the results of these architectural improvements will be in practice, because AMD has not yet made anything concrete about actual cards other than these details.