RDNA 2 vs Nvidia Ampere: Architecture, Cache, and Game Performance

Architecture Overview

AMD switched from the long‑standing GCN design to RDNA in 2019; RDNA 2 builds on RDNA 1 by adding ray‑tracing support and several performance enhancements. Each Work‑Group Processor (WGP) contains four SIMD units, each SIMD has a 32‑wide execution engine, a 128 KB vector register file and can track up to 16 wavefronts (down from 20 in RDNA 1).

New Instructions and Ray‑Tracing

RDNA 2 introduces dot‑product instructions such as V_DOT2_F32_F16 to accelerate machine‑learning workloads and adds hardware‑accelerated ray‑tracing via texture‑unit instructions that perform BVH box tests and triangle tests. The texture units now execute intersection tests instead of traditional texture sampling.

Cache Hierarchy

RDNA 2 retains the WGP core but adds a four‑level cache system: 128 KB L0, 128 KB L1, a 4 MB L2 shared across the die, and a 128 MB “Infinity Cache” (named MALL) that sits between L2 and VRAM. The Infinity Cache captures most memory traffic, dramatically reducing external bandwidth demand, especially at larger test scales.

Latency and Bandwidth

Latency tests show RDNA 2’s multi‑level cache delivers lower latency than Nvidia Ampere when the workload exceeds Ampere’s L2 capacity. Bandwidth measurements indicate that a single WGP can achieve very high cache bandwidth at high clock speeds, and the Infinity Cache scales bandwidth effectively for large data sets.

CU and WGP Modes

RDNA 2 can operate in WGP mode (128 KB shared LDS) or CU mode (two 64 KB LDS halves). LDS latency is ~19.5 ns in both modes. CU‑level memory pipelines and 16 KB L0 vector caches enable hardware‑accelerated ray‑tracing and improve cache utilization.

Game‑Level Analysis

Using Radeon GPU Profiler, the article examines Cyberpunk 2077 (RT on/off), a megademo of Titanic Honor & Glory, and the tank simulator Gunner HEAT PC. With ray‑tracing enabled, the RX 6900 XT spends ~21 ms per frame on RT work, executing 5.8 × 10⁸ box‑intersection tests and 1.1 × 10⁸ triangle tests to reach 25.9 FPS. Compute shaders dominate the workload, and vector‑register‑file limits constrain wavefront occupancy, leading to ~51 % vector ALU utilization in the longest RT dispatch.

Conclusions

RDNA 2 marks a turning point for AMD, delivering performance comparable to Nvidia’s Ampere while using less power, largely thanks to its extensive cache hierarchy and modest clock‑speed gains. The architecture’s hardware‑accelerated ray‑tracing and strong compute capabilities position it well for future GPU generations.