2026-06-08
Modern x86 CPUs run AVX-512 (and to a lesser extent AVX2) at a lower clock frequency than scalar code. This isn't a bug — it's a thermal and electrical bargain the CPU makes with itself. Wide vector units burn enormous current when active; sustaining nominal frequency through a 512-bit FMA every cycle would exceed the package's voltage regulator limits and the die's thermal envelope.
Intel formalized this with frequency licenses. On Skylake-X / Cascade Lake, a core operates in one of three states:
The transitions aren't instant. When a core hits an AVX-512 instruction, it requests a higher license. The voltage regulator ramps up, but until it does, the core runs at a guaranteed-safe lower frequency for ~20µs. After the AVX-512 burst ends, it stays in the elevated license for ~2ms before dropping back — because thrashing licenses would cost more than just staying there.
Real-world example: A web server that occasionally calls memcpy compiled with AVX-512 can throttle the entire core for milliseconds after each call. If your hot path is scalar pointer-chasing and you sprinkle in a single AVX-512 routine, the scalar code runs slower for the next ~2ms than it would have if you'd used AVX2. Cloudflare famously disabled AVX-512 in their workloads for exactly this reason around 2017-2018.
Rule of thumb: AVX-512 is a net win only if the vectorized region runs long enough to amortize both the frequency transition (~20µs) and the trailing penalty on neighboring scalar code (~2ms). For a 4 GHz → 3 GHz drop (25%), you need the vector region to be at least ~4× faster than scalar to break even on its own runtime — and that's before accounting for the tail penalty on whatever runs next.
Ice Lake and later mostly fixed this for client chips by improving voltage regulator response and per-instruction power gating, but server SKUs (Sapphire Rapids, Emerald Rapids) still throttle measurably under sustained AVX-512 load. AMD's Zen 4 handles AVX-512 by double-pumping 256-bit units instead of widening them — no license states, no thermal cliff, but also no peak throughput advantage.