Deep microarchitectural teardown of Geekbench 6 workloads using Intel SDE, PMCs, and SPEC CPU2017 comparisons across Lion Cove, Zen 5, and Neoverse cores.
Key Takeaways
Geekbench 6 is heavily vectorized: AVX2 dominates most workloads; AVX-512 is critical for Background Blur, Object Detection, and Structure from Motion on Granite Rapids targets.
AMX accounts for under 0.2% of instructions in Object Detection but dramatically reduces AVX-512 instruction count, showing outsized impact from sparse use.
IPC distribution is tighter and higher than SPEC CPU2017’s integer suite; Navigation is the lone persistently low-IPC outlier due to extreme branch misprediction (high MPKI).
Lion Cove’s 3 MB L2 and 192 KB L1.5D absorb most L1D misses; Object Remover is the only test with significant L2 miss traffic.
Photo Filter is the most DRAM-bandwidth-intensive Geekbench 6 test despite low L3 MPKI, indicating aggressive prefetcher activity masking latency.
