AMD Shows Off Some Aspects of Vega GPUs at CES 2017
Yesterday AMD showed off some more information about its upcoming Vega GPU architecture. While many were hoping for a paper launch or something telling us when we will get our hands on it, there was still some interesting information presented. Some of the information was already known from previous announcements, such as the use of HBM2 and its speeds, and from leaks, like the improvements to the compute engines, but there is feature one I at least do not recall coming up before and it has to do with the cache.
I am not an expert in how graphics cards work but part of their operation involves loading the various graphical assets of a game into the local VRAM, so the GPU can rapidly call on it for one purpose or another. The results of this processing, such as the finished frame, will then be stored in the VRAM. Vega will change this up some by apparently abstracting this cache system using the High-Bandwidth Controller, as AMD calls it. This controller can use the local High Bandwidth Cache, like HBM2 or possibly GDDR5 if that is present, but it can also use system RAM, non-volatile RAM such as SSDs, and based on what it shows in the video, even network storage. That last location I am guessing is more for professional compute applications than gamers, but it is still stated that Vega will have the capability to work with assets that are streamed in directly from these other caches, without having to wait for it to be placed in the local cache. In total the controller can build a 512 TB virtual address space it can pull from, which is a lot to manage, especially with different memory technologies having different speeds and latencies. Selecting where information is stored, in the local cache or in another location, can also be decided automatically instead of putting on the developers to decide. We will have to wait for reviews to see how well this performs.
Along with this new approach to the cache/memory system, Vega will also use a new geometry pipeline, enabling more than double throughput per clock, and a new computer engine that supports higher clock speeds and has a better IPC than previous AMD GPUs. Vega is still expected to launch H1 2017.