AMD Linux Drivers Detail Vega 10 Core Configuration

Samuel Wan / 7 years ago

As we near launch in Q2, more and more AMD RX Vega GPU details have started leaking. one of the best sources of information are the pre-release drivers, especially for Linux which are publically available. As part of a Direct Rendering Manager update to their Linux driver, we now have some details for the upcoming GPUs. From the drivers, it appears that Vega shares a lot of similarities with their previous Fiji flagship.

Next-Generation Compute Unit!

Just like Fiji, we have 64 CUs but these have been updated to the Next-Generation Compute Unit. This gives us 4,096 shaders based on the new GCN architecture, divided into a total of 4 shader engines. Each shader engine features 2 Asynchronous Compute Units, 16 ROPs (64 total) and 64 TMUs (256 total). This setup allows for 8 different independent workloads threads to be scheduled and handled at once.

_{adev->gfx.config.max_shader_engines = 4;}
_{adev->gfx.config.max_tile_pipes = 8;}
_{adev->gfx.config.max_cu_per_sh = 16;}
_{adev->gfx.config.max_sh_per_se = 1;}
_{adev->gfx.config.max_backends_per_se = 4;}
_{adev->gfx.config.max_texture_channel_caches = 16;}
_{adev->gfx.config.max_gprs = 256;}
_{adev->gfx.config.max_gs_threads = 32;}
_{adev->gfx.config.max_hw_contexts = 8;}

Vega uses HBM2

Unlike Fiji which featured 4GB of HBM in 4 blocks over a 4,096-bit bus, Vega uses HBM2. Due to the increased density and speeds, Vega 10 is expected to get away with just 2 HBM2 blocks over a 2,048-bit bus. Along with what is expected to be a vastly higher clock speed and improved architecture, Vega should leave Fiji in the dust.