Radeon RX 580
The Polaris 20 GPU is an updated version of the Polaris 10 core found in the Radeon RX 480. As a result, the composition of the computing units has not changed. Like Polaris 10, the fully functional version of Polaris 20 has 2304 shader ALUs, 144 texture mapping units and 32 ROPs, and the chip is connected to the RAM via a 256-bit bus with a bandwidth of 8 Gbps per pin.
Let us recall that from the point of view of microarchitecture, the Polaris series has a number of differences from previous AMD products based on GCN logic version 1.2 (the Fiji chip in the Radeon R9 Nano, R9 Fury and Fury X), among which an important place is occupied by color compression with a ratio of up to 8 :1, allowing more efficient use of RAM bus bandwidth. Among other optimizations of the internal structure of the GPU presented in GCN 1.4, we note the optimization of the geometry engine, which is capable of cutting off polygons of zero size or those without pixels in the projection at the early stages of the pipeline. The Compute Unit (the main GCN structure that combines shader ALUs and texture mapping units) has also been tuned in Polaris to improve its specific performance.
Functional innovations in Polaris include support for the half-bit number formats int16 and FP16, which are used in computational tasks that do not require more precise representation (such as machine learning).
The dedicated GPU codec allows hardware to encode and decode video in HEVC Main 10 and VP9 formats with resolutions up to 4K. The display controller supports signal output via DisplayPort 1.3/1.4 and HDMI 2.0b interfaces, as well as high dynamic range (HDR).
Radeon RX 580 Specifications
|
|
||||||
|
Chip
|
||||||
|
Frequencies
|
||||||
|
Memory
|
||||||
|
Interface and TDP
|
AMD set the base and boost frequencies to 1.257 and 1.340 MHz. Partner video cards will have even higher clock speeds, they can exceed 1.400 MHz. The Radeon RX 580 video card is equipped with 4 or 8 GB of video memory. But we recommend taking the 8 GB option so that performance is not limited by the amount of memory. Video memory is connected via a 256-bit interface.
The clock frequency is 2.000 MHz, so the throughput reaches 256 GB/s. We note 36 CUs, 2.304 stream processors, 144 texture units and 32 raster operations pipelines (ROPs).