Get the latest tech news
Disabling Zen 5's Op Cache and Exploring Its Clustered Decoder
Zen 5 has an interesting frontend setup with a pair of fetch and decode clusters.
Multiplying thread IPC by 2 because I’m running SPEC rate tests with two copies pinned to SMT siblings, and the core should spend negligible time in ST mode Performance counters even suggest 525.x264 gains a bit of IPC in decoder-only mode, though that didn’t translate into a score advantage likely due to varying boost clocks. And I still think the graph above is a cool illustration of how the 4-wide decoder starts running out of steam during higher IPC sections of the benchmark, while the op cache has plenty of throughput left on tap.
Or read this on Hacker News