AMD has announced the release of AOCL 6.0.0, a major update to its AMD Optimizing CPU Libraries, tailored specifically for the newly launched 6th Gen AMD EPYCâ„¢ server CPUs (code-named "Venice") built on the Zen 6 architecture. The update brings significant performance improvements across AI inference, numerical computing, memory operations, and data analytics, leveraging Venice-specific optimizations.

Performance Gains Across Key Workloads

Early testing of AOCL 6.0.0 has demonstrated substantial performance gains compared to the previous version, AOCL 5.3:

  • Up to 1.84x faster per-core INT8 GEMM performance for AI inference.
  • Up to 1.5x higher performance for selected mathematical functions.
  • Approximately 15% improvement in memory-intensive operations.
  • 1.3x higher analytics performance per core, with up to 4x acceleration for selected nearest-neighbor analytics workloads.

These improvements are attributed to enhanced utilization of cores, caches, memory bandwidth, and vector units in the Venice processors.

AI Inference and Memory Operations

AOCL-DLP, the library for AI inference, has been optimized for INT8 and FP16 compute paths on Venice processors. This results in up to 1.84x faster per-core INT8 GEMM performance, combining hardware and software gains. Additionally, AOCL-LibMem has improved key memory operations, including up to 30% faster memcpy and up to 20% faster memset operations, which are critical for databases, analytics platforms, and AI frameworks.

HPC and Scientific Computing

For high-performance computing (HPC) workloads, AOCL-BLAS has improved DGEMM efficiency from 95.5% to 97.3%, alongside up to 1.5x acceleration for selected mathematical functions. These updates enable existing applications to benefit from performance gains without requiring source-code changes.

Data Analytics Enhancements

AOCL-DA, the library for data analytics, introduces new APIs for nonnegative matrix factorization, kernel PCA, and t-SNE, as well as optimizations for linear models, SVMs, and decision trees. On Venice systems, per-core performance has improved by approximately 1.3x, with selected ANN workloads achieving up to 4x speedups.

Introduction of AOCL-OpenRNG

AOCL 6.0.0 also debuts AOCL-OpenRNG, an open-source random number generation library optimized for AMD processors. It supports 11 base generators, 17 statistical distributions, and AVX2-optimized kernels, catering to workloads in simulation, financial modeling, and AI.

With these updates, AOCL 6.0.0 positions AMD's 6th Gen EPYC processors as a robust choice for AI, analytics, and HPC applications, offering enhanced performance and efficiency for modern workloads.