Loading the document
1 free read left. Create a free account for unlimited reading + AI research.
Create free accountPresentation
AMD
This AMD presentation (its text references availability on July 7th) describes the RDNA GPU architecture and compares it with the earlier GCN architecture, using the RX 5700 XT (RDNA) and RX Vega 64 (GCN) as reference products. Its scope covers the Workgroup Processor (WGP) structure, instruction issue, wave sizes, register files, caches, load/store queues and memory hierarchy. The document presents RDNA as an architecture focused on graphics and latency, in contrast to GCN's compute and throughput focus. It states that each WGP provides resources of two Compute Units to a single workgroup and that Navi has 20 WGPs, corresponding to 40 CUs. On issue, it states that RDNA issues one instruction every cycle per SIMD32, where GCN issued one every four cycles per SIMD16. It also states that 256 threads per WGP often reach over 90% ALU utilization in practice, citing observations from pixel shaders in a game rendering trace. RDNA features 128B cache lines and a new 128KB L1 cache per shader array.
Each row is a game or company this document names, each column a document, oldest first. A mark means that document names it. Select a row to see those documents.
AMD · 2025
Comments