AOCL 6.0.0 Tuned for 6th Gen AMD EPYC™ CPUs: What’s New for HPC, AI, and Analytics

Nov 19, 2026

Shot of Corridor in Working Data Center Full of Rack Servers and Supercomputers with High Speed Inernet Visualization Projection.

Introduction

AMD recently announced the general availability of 6th Gen AMD EPYC™ Server CPUs (code-named "Venice"), built on the new "Zen 6" architecture. A new processor is only as fast as the software running on it, so with AOCL 6.0.0 we have tuned our libraries for these processors so your applications can make better use of the hardware.

AMD Optimizing CPU Libraries (AOCL) 6.0.0 includes Venice-specific optimizations across numerical computing, memory operations, AI inference, data analytics, compression, and runtime infrastructure. In this post, we will walk through where those changes made a measurable difference in early testing.

Here's a quick look at the highlights across AI inference, math libraries, memory operations, and data analytics. Gains vary by workload, and we cover each area in more detail in the sections that follow.

Early testing shows the following gains over 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
  • Approximately 1.3x higher analytics performance per core
  • Up to 4x acceleration for selected nearest-neighbor analytics workloads

Unless we say otherwise, the improvements in this post compare AOCL 6.0.0 with the previous release, AOCL 5.3, using representative workloads and benchmark configurations.

Venice Support in AOCL 6.0.0

In AOCL 6.0.0, we've added Venice-specific optimizations to selected routines and workloads, and we plan to do more in later releases.

Performance comes down to how well software uses cores, caches, memory bandwidth, and vector units. For this release, we focused on the areas where our benchmarking showed the most room for improvement: AI inference, numerical computing, memory-intensive operations, and data analytics.

As we gather new workload patterns and benchmarking data, we will extend the Venice-specific optimizations to more libraries.

AI Inference (AOCL-DLP)

More and more recommendation systems, retrieval pipelines, ranking models, and enterprise AI services are running inference on CPUs. AOCL-DLP provides optimized low-precision compute primitives for these workloads.

In AOCL 6.0.0, AOCL-DLP improves selected INT8 and FP16 compute paths on Venice, with further optimizations planned for future releases.

Here's how per-core INT8 GEMM speedup compares with 5th Gen AMD EPYC™ Server CPUs  (code-named "Turin"):

Configuration

INT8 GEMM speedup (per-core)

Venice hardware alone

1.4x–1.7x

Venice with AOCL 6.0.0 optimizations

1.6x–1.84x

Per-core, AOCL's software optimizations add up to a further 12% on top of the hardware gain. The result is higher inference throughput, and you keep the deployment model that CPUs already offer.

Memory Operations (AOCL-LibMem)

Many applications spend as much time moving data as computing on it. Databases, analytics platforms, AI frameworks, storage systems, and cloud services all rely on fast memory operations.

AOCL-LibMem 6.0.0 improves key memory operations and adds tuning for AMD EPYC Venice processors. Testing also included workloads from DCPerf, Meta’s open-source benchmark suite. Here's how that looks:

Workload

Improvement

Geomean memory performance

~15%

Hot-cache memcpy

Up to 30%

memset operations

Up to 20%

DCPerf tensor workloads

6-7%

Many applications call memcpy and memset heavily, so even these gains can show up as end-to-end improvements.

HPC and Scientific Computing

HPC workloads lean heavily on dense linear algebra and mathematical functions. AOCL-BLAS library provides the dense matrix kernels, such as DGEMM, that many of these applications depend on.

In AOCL 6.0.0, we've improved AOCL-BLAS DGEMM along with selected math routines that frequently affect overall application performance.

Here's what early testing showed:

  • AOCL-BLAS DGEMM efficiency improved from 95.5% to 97.3%
  • Up to 3% improvement in HPL performance
  • Mathematical routines delivering 1.25x average performance improvements
  • Up to 1.5x acceleration for selected functions

Existing applications can pick up these gains without any source-code changes.

Data Analytics (AOCL-DA)

Analytics workloads work through large datasets and feed recommendation systems and machine learning workflows. AOCL Data Analytics (AOCL-DA) library provides optimized algorithms for these workloads, and the 6.0.0 release adds new APIs and Zen 6 tuning.

Here's what's new in AOCL-DA 6.0.0:

  • New nonnegative matrix factorization, kernel PCA and t-SNE APIs
  • Scalar quantization for ANN, with a FAISS patch for easy access to optimized ANN algorithms
  • SciPy patch for optimized basic statistics, interpolation and nonlinear least squares
  • Faster linear models, SVM, ANN, decision trees and metrics
  • Zen 6-optimized code paths, including internal half-precision use

On Venice systems, we saw per-core performance improve by approximately 1.3x, and selected ANN workloads ran up to 4x faster. Put together, these changes mean shorter run times for analytics jobs on large datasets.

Want a practical example? The AMD technical article Scikit Workloads with AOCL Data Analytics shows how AOCL-DA can improve performance for common machine learning and analytics pipelines that use Scikit-Learn.

Other Changes in AOCL 6.0.0

This post covers selected Venice-related changes, but AOCL 6.0.0 includes more: updates across numerical computing, AI, analytics, compression, memory optimization, random number generation, and developer infrastructure.

New Library: AOCL-OpenRNG

AOCL 6.0.0 introduces AOCL-OpenRNG, a new open-source random number generation library for simulation, financial modeling, scientific computing, AI, and analytics workloads. It complements AOCL-RNG, builds on Arm's upstream OpenRNG implementation, and remains compatible with Arm processors. Its Vector Statistical Library (VSL)-compatible APIs make migration easier for applications already using that interface, and its open-source implementation allows integration into open-source software.

AOCL-OpenRNG also includes AMD-specific performance enhancements, architecture-aware optimizations, and a modern build infrastructure, and we are developing it with community input. You can choose between AOCL-RNG and AOCL-OpenRNG based on your deployment, portability, and open-source requirements.

Other libraries that got updates in this release:

  • AOCL-LibM, with improved mathematical routine performance and expanded support for emerging AMD architectures. AOCL-LibM is also used by the AOCC compiler, so these improvements carry through to compiled applications, including SPEC CPU® 2026 workloads that call math library functions.
  • AOCL-BLAS and LAPACK, delivering continued improvements for scientific computing and numerical applications.
  • AOCL-Compression, including enhancements driven by modern data-processing and cloud workloads.
  • AOCL-FFTZ, with ongoing performance and usability improvements for FFT workloads.

Looking Ahead

With AOCL 6.0.0, applications can make better use of 6th Gen AMD EPYC™ Server CPUs across AI, scientific computing, analytics, and enterprise workloads.

The optimizations are architecture-aware, and the APIs stay familiar, so adoption is straightforward.

We will keep adding architecture-aware optimizations as Venice platforms mature. In upcoming technical blogs, we will dig into individual library changes, optimization techniques, and workload-specific results from AOCL 6.0.0.

Ready to try it? Explore AOCL 6.0.0 at the AOCL Developer portal, and stay tuned for detailed insights from our engineering team.

Footnotes

Disclaimer:

System Configuration

  • AMD EPYC Venice System Model: AMD Reference Platform
  • CPU: 2x AMD EPYC 9996 256-Core Processor (512 cores, 512 threads; SMT disabled)
  • NUMA: 4 NUMA nodes, 2 per socket. NUMA auto-balancing enabled
  • Memory: 4096 GiB (32 x 128 GiB dual-rank x4 DDR5 RDIMMs; 4026 GiB OS-visible)
  • Disk: 1x 1 TB Samsung NVMe SSD (1.0 TB / 953 GiB raw)
  • Host OS: Ubuntu 26.04 LTS
  • Kernel: 7.0.0-34-generic
  • System BIOS: WNIV6325N, dated 2026-03-25
  • System BIOS Vendor: AMD Corporation

Performance results are based on AMD internal testing using representative workloads. Actual results may vary based on hardware configuration, software environment, compiler settings, workload characteristics, and other factors. Reported improvements reflect selected workloads and are not necessarily representative of all applications. 

 

Share:

Article By


PMTS Software System Design Eng.

Related Blogs