RC RANDOM CHAOS

Zen 6: AMD's First Split Between Classic and Dense Cores

· via Hacker News

Original source

AMD's Zen 6 Family: Feeding the Beast

Hacker News →

Zen 6 features substantial architectural redesign compared to Zen 5’s iterative improvements, with major changes to branch handling, decode logic, and out-of-order resources. Branch misprediction recovery latency drops from 3.5–4ns to 3–3.5ns per cycle. Single threads can now utilize both decode clusters. Integer throughput improves to 5.96 64-bit adds per cycle despite a reduction in multiply units, and VNNI capacity doubles from two to four units for AI inference.

For the first time, AMD’s Classic and Dense cores diverge architecturally. Zen 6c significantly expands out-of-order structures to handle higher memory latency in 32-core CCDs: reorder buffers grow by 64 entries, integer register files by 64, vector schedulers by 42, and load queues by 50. Zen 6 targets high frequency in 12-core configurations. L3 cache latency increases to 55–66 cycles from Zen 5’s 47 due to larger caches and expanded mesh topology, though L2 bandwidth improves to a consistent 64 bytes per cycle.

The EPYC 9686F (Zen 6) leads the EPYC 9996 (Zen 6c) by roughly 18% in SPEC CPU 2026 benchmarks. However, Zen 6c competes effectively in specific integer workloads where higher clock speed provides minimal advantage.

Read the full article

Continue reading at Hacker News →

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