Scaling & Optimizing Performance on AMD EPYC Systems
summary
Focuses on concrete tuning: the 9000x series uses a single IO die with up to 12 memory channels and configurable NPS settings; NPS4 lowest per‑core latency by pinning cores to quadrants. Benchmarks (Feronx) show a ~2% geometric‑mean gain overall but 10–15%+ on HPC/memory‑bound workloads. One‑DPC vs two‑DPC drops DDR speed (to ~4000MT/s) and can hurt throughput; a memory‑population matrix and AMD docs/tools (tech portal, SEV/SNP/TSME guides) and practical tuning steps are provided.
AI-written summary · about
chapters (7)
- 0:00Introduction and AMD overview
- 3:31AMD EPYC architecture and scalability
- 7:01Memory subsystem and NPS settings
- 12:27Memory channel population and performance
- 19:56Memory sizing and optimization
- 24:13Confidential computing overview
- 28:48Resources and support
related talks
45mUsing hardware hints for optimal page placement
30mEncrypting Memory at Scale
19mDo nothing fast: How to scale idle cpus?
01h 11mPerformance engineering - being friendly to your hardware
24mOptimizing Storage Performance for 4–5 Million IOPs
43mLive in a world with multiple memory types
22mKernel-Native Packet Processing on AMD GPUs: XDP Offload without Userspace Runtimes
43mExpanding the Memory Limits with CXL Software Stack