Compute Unit RDNA 2 contains a significant increase in the number of Compute Units (CUs) with a maximum of 80, a doubling from the maximum of 40 in the Radeon RX 5700 XT. CUs are organized into groups of two named Work Group Processors with 32KB of shared L0 cache per WGP. Each CU contains two sets of an SIMD32 vector unit, an SISD scalar unit, textures units, and a stack of various caches. New low precision data types like INT4 and INT8 are new supported data types for RDNA 2 CUs.
Ray tracing Real-time hardware accelerated
ray tracing is a new feature for RDNA 2 which is handled by a dedicated ray accelerator inside each CU. Ray tracing on RDNA 2 relies on the more open DirectX Raytracing protocol rather than the
Nvidia RTX protocol. In February 2023, it was reported that driver updates had boosted ray tracing performance by up to 40% using DirectX Raytracing.
Clock speeds With RDNA 2 using the same 7 nm node as RDNA, AMD claims that RDNA 2 achieves a 30% frequency increase over its predecessor while using the same power.
Cache and memory subsystem In addition to the traditional L1 and L2 caches that GPUs possess, RDNA 2 adds a new global L3 cache that AMD calls "Infinity Cache". This was done to avoid the use of a wider memory bus while still being able to maintain the same data bandwidth. Product technology architect
Sam Naffziger said that, without Infinity Cache, "We were looking at the daunting prospect of having to put a 512-bit interface and all the power, area and expense associated with that". Using a wider memory bus requires more power which is in conflict with AMD's increased performance-per-watt goals for RDNA 2. AMD engineers ran tests comparing RDNA 2 silicon featuring a large on-die cache and with wider memory buses. They discovered that having such a cache would aid in the re-use of temporal and spatial data when the GPU is rendering a complex image. It is beneficial for the GPU's compute units to have fast access to a physically close cache rather than searching for data in video memory. AMD claims that RDNA 2's 128MB of on-die Infinity Cache "dramatically reduces latency and power consumption". The GPU having access to a large L2 or L3 cache allows it to more quickly access necessary data compared to accessing VRAM or system RAM. The Infinity Cache is made up of two sets of 64MB cache that can run on its own clock rate independent from the GPU cores. The Infinity Cache has a peak internal transfer bandwidth of 1986.6 GB/s and results in less reliance being placed on the GPU's
GDDR6 memory controllers.
Media engine RDNA 2 uses the
VCN 3.0, VCN 3.1, and VCN 3.1.2 video decoding blocks in its media engine. It adds support for
AV1 decoding at up to
8K resolution, though AV1 hardware encoding support would not come until
RDNA 3 in 2022. However, the low-end Navi 24 die and iGPUs based on RDNA 2.0 do not contain any media encoders and cannot decode AV1 as a result. == Navi 2x dies ==