How AMD's Infinity Cache Solved a Massive Hardware Dilemma

Science · Published 1 November 2026 Swipe more like this →

In short: To boost performance, engineers often face a brutal trade-off: wider memory buses provide speed but consume massive amounts of power and physical space. AMD broke this cycle with RDNA 2 by moving the data closer to the processor, using a unique on-die cache to bypass traditional hardware bottlenecks.

Released in November 2020, the RDNA 2 microarchitecture was designed to prioritize power efficiency and performance-per-watt. While its predecessor, RDNA, focused on replacing the aging Graphics Core Next (GCN) architecture, RDNA 2 aimed for a much more aggressive leap. AMD engineers targeted a 50% uplift in performance-per-watt by reconfiguring the graphics pipeline to move caches closer to shader engines and introducing mesh shaders, which allow rendering to be done in parallel using smaller 'meshlets'.

The most significant architectural shift was the introduction of 'Infinity Cache.' To maintain high data bandwidth without the 'daunting prospect' of a 512-bit memory interface—which would have been expensive and power-hungry—AMD implemented a large on-die L3 cache. Operating at separate clock speeds, this storage unit comprises two 64 MB blocks to achieve a maximum internal transfer rate of 1986.6 GB/s. By keeping temporal and spatial data physically close to the compute units, the GPU avoids the latency and power costs of constantly searching for data in external video memory (VRAM).

The results were measurable: RDNA 2 achieved a 30% frequency increase over the original RDNA while maintaining the same power profile. It also delivered up to a 54% increase in performance-per-watt, with roughly 21% of that gain attributed specifically to performance-per-clock improvements fueled by the Infinity Cache. This architecture now powers major hardware, including the PlayStation 5 and Xbox Series X/S consoles, as well as the Steam Deck.

Beyond memory management, RDNA 2 introduced dedicated ray accelerators within each Compute Unit (CU) for real-time hardware accelerated ray tracing. Unlike Nvidia's proprietary protocols, AMD utilized the more open DirectX Raytracing protocol. While early drivers were a factor, subsequent updates in February 2023 reportedly boosted these specific performance metrics by up to 40%.

General information, not professional advice. Spotted an error? Flag the card or see corrections and complaints.

Intellect Plus is a visual educational discovery platform for exploring fascinating facts and ideas. Brain Food, Not Brain Rot. What is Intellect Plus?

Related

• How AMD's RDNA architecture shrunk the 'wavefront' to boost gaming power
• How AMD’s Zen 3 architecture redefined desktop performance through a radical core redesign
• How GPUs hide memory latency by juggling 64 threads at once
• Why your computer stopped getting faster and started getting more cores instead
• The software that dominated 95% of the professional publishing market
• The font that powers half of Google's web views
• The fortress of privacy faces a paradox of trust
• How a backported microarchitecture redefined the limits of 14nm silicon
• The engineering required to sustain 1600 nits of brightness indefinitely
• The tiny annotations that bridge the gap between logographs and speech
• The architecture of choice for high-end servers and embedded systems
• The browser that prioritizes one-handed navigation and ad-blocking over defaults

More in Found by OrchestratorBot · More in Science · Latest · A–Z · All topics