The basis of 3D-card AMD Radeon RX Vega Liquid Cooled Edition 64 RX 64 RX Vega and Vega 56 was the graphic architecture of the new generation
Continuing the story of 3D-card AMD Radeon RX Vega Liquid Cooled Edition 64 RX 64 RX Vega and Vega 56, it is necessary to stay on the GPU, which is its basis.
It is built on a new architecture Vega, consists of 12.5 billion transistors and is produced on 14-nanometer technology FinFET LPP. The processor is designed for operation at a clock frequency up to 1.7 GHz and includes about 45 MB of SRAM, and for inner joins, it uses bus Infinity Fabric.
The processor has four asynchronous engine executes the commands, four modified geometric engine, and sixty-four compute units Compute Unit (CU) of the new generation. According to AMD, they seriously reworked compared with the previous architectures. The CU is 4096 stream processors, 256 texture units and 64 ROP unit.
When designing GPUs, AMD engineers have used experience architecting Zen, in particular, shortened and optimized all the lines of data in critical areas, thus reducing the latency by 8%, the crystal size by 18% while the energy consumption is almost doubled.
Other changes: note doubled to 4 MB cache second level, which now can use blocks ROP, which is important for techniques that use rendering into textures. Also the chip is using high-speed memory HBM second generation ECC, important for professional applications, featuring higher density (up to 8 times more memory on the stack) and bandwidth. HBM2 has a 2048 bit bus and operates at a frequency close to 1 GHz, which gives a throughput of 484 GB/s.
AMD believe that the memory HBM should be viewed as a cache for data that reside on slow media. In fact, HBM2 in Vega is used as a cache of the last level for system memory and data storage. The virtual memory architecture based on pages of different sizes, dormant of which are moved to slower storage, freeing the cache for active memory pages. It does not require complicated memory management, and store all data in fast local memory (cache) is not required. This approach allows you to use a virtual endless amount of resources of many millions of triangles and hundreds of gigabytes of texture data.
Compute units Compute Unit is now support double the pace for the operations on the data with precision FP16, compatible with IEEE. Measurements of AMD using FP16 or mixed precision compared to the precision FP32 gives in 3DMark Sierra during post-processing of approximately 20-25% advantage with no visible loss in quality. It can already be used in existing graphics APIs.
In addition, Vega has added 40 new instructions: combined operations to increase the number of executable operations per clock, for use in the mining of cryptocurrencies, for processing video data and memory addressing. And graphic changes note the full DirectX Feature Level 12_1, which exceeds not only the AMD GPU of last generation, but the best NVIDIA — Pascal.
In particular, there is support for Raster Ordered Views, Conservative Rasterization Tier 3 and Tiled Resources Tier 3, which is better than the GPU of a competitor. An additional level of Conservative Rasterization allows the use of these features in algorithms for efficient discard invisible geometry.
In General, AMD has long been calling for changing the existing geometric pipeline and the introduction of new types of Shader Primitive Shaders, a new hardware pipeline stage, which will allow to discard invisible primitives very early, increasing overall efficiency.
For example, while Vega itself is twice as fast as Fiji for processing geometry, the proposed change of the conveyor allows to increase the speed even tripled. Alas, that it is impossible to use the current API, but the company cooperates with leading developers in order to get them interested.
Will not go away effective use of asynchronous computing, which is an advantage of AMD solutions. The use of this feature in DX12 version of The Division gave an increase of 13%, and the specific optimization of the basic process of rendering DOOM on the PS4 ported to PC, and all gave 43% increase in speed. Has been improved tile optimization in Draw Stream Binning Rasterizer, which gives the average gains of 10-15% (sometimes more) in different software (not just games), and improved geometric engines for loads with a large amount of geometry.
All this and much more we’ll have more details in the review of Radeon RX Vega.