The AI Compute Extensions (ACE) for x86 [pdf] (x86ecosystem.org)

The AI Compute Extensions (ACE) for x86 were recently announced in a whitepaper co-authored by teams from AMD and Intel, detailing a new architecture designed to enhance the performance, scalability, and energy efficiency of matrix multiplication—crucial for machine learning and neural networks. ACE integrates seamlessly with AVX10, allowing for substantial improvements in computational density while utilizing existing AVX infrastructure. The outer product operations introduced by ACE promise a 16× performance gain over traditional AVX10 multiply-accumulate operations without significantly increasing processing latency. This development is significant for the AI/ML community as it democratizes AI capabilities across platforms—from laptops to data centers—by reducing the friction associated with offloading computational tasks to specialized hardware. ACE supports a variety of data formats crucial for machine learning applications, including inline block scaling and low-precision types like INT8 and BF16, providing flexible performance tuning for diverse implementations. Additionally, the introduction of efficient format conversion instructions enhances the usability of ACE in future-proofing software and adapting to evolving data storage needs in AI applications.