Quantifying performance bottlenecks and jagged benchmark results
In stress test 2026-AMD-09, with ambient temp locked at 22℃, I ran 3DMark loops. The scores were swinging wildly between 18,000 - 21,000, which made me think my EXPO profile wasn't kicking in. To dig deeper, I ran five consecutive passes and used Cinebench to track per-core clock jitter, discovering the bottleneck was actually L3 cache scheduling latency. I tweaked the Precision Boost Overdrive curve, setting a negative voltage offset of -0.050V, which tightened the score variance to within 2,000 points. Performance is up, but in a few extreme scenarios, I still get a sudden frame drop due to core frequency switching. It's not a perfect straight line, but it's close.
