x86-64-v4 is AVX-512: 512-bit vector registers (32 of them), mask registers for per-element predication, and the subsets for byte and word elements (BW), doublewords and quadwords (DQ), conflict detection (CD) and the same instructions on 128- and 256-bit vectors (VL).

Features added by x86-64-v4, on top of all lower levels.
FeatureWhat it providesEnumerated byLinux flagExample
AVX512FAVX-512 foundation.CPUID.(EAX=07H,ECX=0):EBX[16]avx512fkmovw
AVX512BWAVX-512 byte and word instructions.CPUID.(EAX=07H,ECX=0):EBX[30]avx512bwvdbpsadbw
AVX512CDAVX-512 conflict detection.CPUID.(EAX=07H,ECX=0):EBX[28]avx512cdvplzcntd
AVX512DQAVX-512 doubleword and quadword instructions.CPUID.(EAX=07H,ECX=0):EBX[17]avx512dqvpmullq
AVX512VLAVX-512 instructions on 128- and 256-bit vectors.CPUID.(EAX=07H,ECX=0):EBX[31]avx512vln/a

Processors at this level

Intel Core (client)
Cannon Lake (2018), Ice Lake (client) (2019), Tiger Lake (2020), Rocket Lake (2021), Nova Lake (announced)
Intel Xeon (server)
Skylake-SP (2017), Cascade Lake (2019), Cooper Lake (2020), Ice Lake (server) (2021), Emerald Rapids (2023), Sapphire Rapids (2023), Granite Rapids (2024), Diamond Rapids (announced)
AMD (high performance)
Zen 4 (2022), Zen 5 (2024), Zen 6 (announced)

Watch out for

  • Intel’s hybrid client processors, from Alder Lake (2021) to Panther Lake (2026), do not support AVX-512 and are x86-64-v3. Intel has confirmed AVX10.2 for Nova Lake, which brings 512-bit vectors back to its client processors.
  • The Xeon Phi processors Knights Landing and Knights Mill implement AVX-512 Foundation and CD but not BW, DQ or VL, so they are x86-64-v3.
  • AMD’s Zen 4 executes 512-bit operations as two 256-bit halves, and Zen 5 has full-width 512-bit units in most models. Both are x86-64-v4: the level is about which instructions exist, not how fast they run.

Building for it

gcc -march=x86-64-v4 ...     # GCC 11+, Clang 12+
GOAMD64=v4 go build ...      # Go 1.18+
RUSTFLAGS="-C target-cpu=x86-64-v4" cargo build ...

The psABI defines no level after x86-64-v4. Later extensions are enabled individually; GCC 15, for example, has -mavx10.2 for AVX10.2 and -mapxf for APX.

Sources

  1. x86-64 psABI: Micro-Architecture Levels
  2. GCC 11 release notes
  3. GCC 15 release notes (AVX10.2, APX)
  4. Intel Officially Confirms AVX10.2 and APX Support in Nova Lake (TechPowerUp)