x86 Ecosystem Advisory Group publishes ACE v2 and AVX10.3 specifications with GCC patches

The x86 Ecosystem Advisory Group, jointly led by AMD and Intel, has published ACE v2 and AVX10.3 architecture specifications for future processors. These extensions build upon existing AVX and ACE capabilities with new low-precision data formats and vector conversion instructions aimed at AI and general-purpose workloads. Intel compiler engineers have already submitted GCC patch series for both instruction sets, continuing the trend of open-source compiler support preceding silicon availability by multiple years.

No processors currently implement ACE or AVX10.2, and hardware with ACE v2 and AVX10.3 is not expected until approximately 2028-2029. The specifications are available through x86ecosystem.org, the group's public specification repository.

ACE v2 and AVX10.3 instruction set changes

• ACE v2 reduced-precision formats: Adds new 4-bit integer formats and expands 8-bit format capabilities for AI inference acceleration, alongside existing FP16, FP32, and FP64 floating-point modes.
• AVX10.3 conversion instructions: Introduces FP8 to BF16 up-convert operations, BF16 to FP8 conversions, and FP16/BF16 to INT4 conversion instructions for mixed-precision vector workloads.
• GCC compiler readiness: Intel has submitted separate patch series to GCC for ACE v2 instructions and AVX10.3 instructions, enabling early toolchain validation before hardware availability.
• Specification availability: Complete technical documentation is posted on x86ecosystem.org, the joint AMD-Intel specification clearinghouse established for collaborative x86 extension development.

Why it matters

Early open-source compiler support for ACE v2 and AVX10.3 ensures that Linux distributions and embedded toolchain maintainers can integrate these extensions well ahead of 2028-2029 hardware launches, avoiding the lag that historically plagued new x86 feature adoption. The AMD-Intel collaboration through the x86 Ecosystem Advisory Group reduces fragmentation risk for developers building AI inference and vectorized applications across both vendors' future processor generations.