RDNA 5
AMD's next-generation GPU architecture, expected to launch in mid-2027 as the successor to RDNA 4.
RDNA 5 is AMD's upcoming consumer GPU architecture, the successor to RDNA 4 and AMD's next swing at competing with Nvidia in both gaming and local AI inference. Credible leaks point to a mid-2027 launch on TSMC's N3P node, though AMD hasn't officially announced specs, pricing, or availability.
Where Things Stand
The architecture has reportedly taped out, which means silicon exists and is being validated, but tape-out is roughly 12-18 months ahead of retail shelves. Everything beyond "N3P node, mid-2027 window" is speculation. AMD has not confirmed VRAM configurations, memory bandwidth targets, compute-unit counts, or ROCm support timelines. Anyone quoting specific tok/s numbers for RDNA 5 today is guessing.
Why Local AI Builders Care
For local LLM work, AMD's relevance hinges on two things: VRAM-per-dollar and software support. RDNA 4 cards undercut Nvidia on raw VRAM pricing but lose ground on the CUDA ecosystem — most inference runtimes target CUDA first, and ROCm support trails by months on new hardware. RDNA 5 will inherit that same software gap unless AMD invests aggressively in day-one ROCm and llama.cpp compatibility. The hardware question is whether AMD pushes VRAM capacity high enough (24GB+ on mid-tier cards) to make waiting worthwhile versus buying a current Nvidia card today.
Why It Matters for Local AI
The "wait for RDNA 5 or buy Nvidia now" question is real for anyone building a rig in 2026. If you need to run 70B models quantized today, waiting 12+ months for unconfirmed specs is a bad trade — depreciation on current Nvidia cards plus a year of lost inference time usually outweighs the upgrade. Wait only if your current setup works and you're optimizing for 2027-2028 model sizes that don't fit in today's VRAM budgets.