RC RANDOM CHAOS

AMD Details Zen 6 Improvements: Denser Cores, AI Focus

· via Hacker News

Original source

Talking Zen 6 with Mike Clark at AMD In-Depth

Hacker News →

AMD’s new Zen 6 architecture introduces two core variants—a high-frequency design and Zen 6c, a denser alternative that sacrifices clock speed for increased instructions per cycle through larger execution queues and complete L3 cache per core (4MB per core across 32 cores for 128MB total). The two variants maintain software compatibility, allowing optimizations to apply uniformly across both designs.

Design improvements span the pipeline. The fetch front-end more efficiently delivers instructions from dual decode blocks with lower power overhead. Integer execution consolidated from three multiply units down to two after empirical testing showed minimal performance gains from the third. The vector unit received two additional VNNI operations supporting AI inference workloads, added late in the design cycle. L2 cache bandwidth improvements now reliably deliver 64 bytes per cycle by reducing prefetch conflicts. The branch predictor improved both misprediction recovery speed and accuracy over Zen 5 through optimized table layouts and faster pipeline redirect mechanisms.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.