Silicon Logo

AMD Accelerated Data Center Premiere

Mastodon Logo

Contents

High Performance Computing

  • Smarter Client Devices & Edge
  • AI & Analytics
  • 5G & Communication Infrastructure
  • Adaptable & Intelligent Systems
  • Gaming, Visualisation & Simulation
  • Cloud, Network, Hyperscalers & Supercomputers

Data Center Workloads

  • Challenging
  • Require Innovation

General Purpose

  • Socket-level Performance

Technical

  • Per-core Performance

Accelerated

  • Highly Parallel
  • Massive Computational Capability

Cloud-Native

  • Maximum Core & Thread Density

Accelerated Computing Leadership

  • Hardware Innovation

  • Software Innovation

  • System Design Innovation

  • Focus on CPUs and GPUs

  • New Cores

  • New Packaging & Process Technologies

  • New Products

  • Combining it all with Exascale Computing

Data Center CPUs

  • Tremendous Progress for the past 4 years
  • New Trajectory for Performance & Scalability
  • New levels of Data Security
  • Growing Customer for Epyc
  • 200 Million EPYC Cores shipped to customers
  • Cloud, Enterprise & HPC

Epyc Hyperscalers

  • Amazon Web Services

  • Alibaba Cloud

  • Baidu Cloud

  • Google Cloud

  • IBM Cloud

  • Oracle Cloud

  • Microsoft Azure

  • Tencent Cloud

  • Twitter

  • Meta (Newest Hyperscaler adopting Epyc)

SAP Collaboration

  • S/4HANA on AMD Epyc
  • Rise with Epyc Offering
  • Reduced TCO$ and Carbon Footprint
  • AMD is one of the first customers

Packaging Leadership

  • 2015: 2.5D HBM
  • 2017: MCM
  • 2019: Chiplets
  • 2021: 3D Chiplets (Chiplets + Advanced 3D Stacking)

Epyc Milan-X

  • 3D Chiplets on Epyc Milan

  • Launch: Q1 2022

  • L3 Cache: 3x Milan

  • L2+L3 Cache/Socket: Up to 804 MiB

  • Cores: Up to 64 Zen 3

  • SP3-Socket Compatible (With BIOS Upgrade)

  • >50% Uplift in targeted Accelerated Computing Workloads

  • Reduces Memory Bandwidth Pressure

  • Reduces Latency [On Average]

  • Take advantage of software without changes

Hybrid Bonding

  • Direct Copper-to-copper Bonding
  • Improves Thermals
  • Increases Interconnect Density
  • 200x Interconnect Density over 2D Chiplets
  • 15x Interconnect Density over Microbump 3D
  • 3x Interconnect Energy Efficiency over Microbump 3D

Technical Computing

  • Finite Element Analysis

  • Structural Analysis

  • Computational Fluid Dynamics

  • Electronic Design Automation

  • Wide ISV Ecosystem (Certified and Highly Performant)

Competition

  • Finite Element Analysis (Ansys Mechanical)

  • 2P Epyc 75F3 has 33% uplift over 2P Xeon 8362

  • Structural Analysis (Altair Radioss)

  • 2P Epyc 75F3 has 34% uplift over 2P Xeon 8362

  • Fluid Dynamics (Ansys CFX)

  • 2P Epyc 75F3 has 40% uplift over 2P Xeon 8362

  • Electronic Design Automation (Synopsys - RDNA 2 Core in 1 hour)

  • 40.6 Jobs on 16C Milan-X over 24.4 Jobs on 16C Milan (66% Uplift)

Data Center Benefits

  • Enterprise Applications
  • Data Mining
  • Risk Analysis
  • Anomaly Detection
  • Media and Entertainment
  • Live Broadcasting
  • Visual Effects
  • Realtime Rendering
  • Artificial Intelligence
  • Recommendation Engines
  • Natural Language Processing
  • Image Reception

Microsoft Azure HPC

  • Azure HBv3 with Milan-X
  • Private Preview - Today
  • GA - Soon
  • 80% Uplift in Ansys Fluent Aerospace Simulations
  • 50% Uplift in Ansys Fluent Automotive Crash Test Modelling

OEM Partners

  • Cisco
  • Dell
  • HP Enterprise
  • Lenovo
  • Supermicro

Instinct Aldeberan (MI200)

  • CDNA 2

  • Designed to enable Exascale Computing (HPC/ML)

  • 3rd Generation Infinity Architecture

  • HPC Performance: 4.9x Nvidia A100

  • AI Performance: 1.2x Nvidia A100

  • Science Leadership (Fueling Exascale)

  • ROCm Open Ecosystem

  • Process Node: TSMC N6

  • Transistors: Up to 58 Billion (2 Dies)

  • Compute Units: 220

  • Matrix Cores: 880

  • Memory: 128 GB HBM2e @ 3.2 TB/s (8 HBM2e Stack)

  • 4x Capacity

  • 2.7x Bandwidth

MI200 Series

  • OAM (Shipping Today)
  • MI250, MI250X
  • PCIe (Coming Soon)
  • MI210

Key Innovations

  • Two CDNA 2 Dies
  • Second Generation Matrix Cores for AI
  • Ultra High Bandwidth Die Interconnect
  • Coherent CPU-to-GPU Interconnect (8 Infinity Fabric Links)
  • 2.5D Elevated Fanout Bridge [ASE FOEB]

Performance in HPC & AI

  • Note: Nvidia A100 vs Instinct MI250X (Advantage)
  • FP64 Vector: 9.7 TF vs 47.9 TF (4.9x)
  • FP32 Vector: 19.5 TF vs 47.9 TF (2.5x)
  • FP64 Matrix: 19.5 TF vs 95.7 TF (4.9x)
  • FP32 Matrix: N/A vs 95.7 TF (N/A)
  • FP16/BF16 Matrix: 312 TF vs 383 TF (1.2x)
  • Memory Capacity: 80 GB vs 128 GB (1.5x)
  • Memory Bandwidth: 2.0 TB/s vs 3.2 TB/s (1.6x)

Application Performance

  • Note: Instinct MI250X vs Nvidia A100

  • AMG: 3.0x

  • HACC: 1.9x

  • HPL: 2.8x

  • LSMS: 1.6x

  • NBODY: 2.5x

  • OpenMM: 2.4x

  • Quicksilver: 2.5x

  • LAMPS ReaxFF (Hydrocarbon Combustion, 20M Atoms)

  • 8.9M Atom Steps/s on Nvidia A100 vs 19.5M Atom Steps/s on Instinct MI250X (2.2x)

3rd Generation Infinity Architecture

  • 800 GB/s Aggregate Bandwidth between CPUs and GPUs
  • CPU-GPU Memory Coherency (Developer Productivity)
  • Accessible Performance

Software (ROCm)

  • Support for major AI frameworks
  • Key ML Models optimised in ROCm 5.0
  • ResNet, BERT, DLRM

Ecosystem

  • Asus
  • Atos
  • Dell
  • Gigabyte
  • HP Enterprise
  • Lenovo
  • Microsoft
  • Supermicro

Frontier (ORNL)

  • Installation in progress
  • Coming online soon
  • Available to scientists in early 2022

Epyc Roadmap

  • Adoption of Milan has outpaced Rome

HPC Process Technology

  • Process Node: N5/N5P HPC
  • Density: 2x
  • Power Efficiency: 2x
  • Performance: 1.25x

Epyc Genoa

  • Process Node: TSMC N5/N5P HPC

  • Cores: 96 Zen 4 Cores

  • Memory: DDR5

  • Interconnects: PCIe 5.0, CXL (incl. CXL.memory)

  • Enhanced Security Features

  • Sampling to Customers

  • On Track for 2022 Production & Launch

Epyc Bergamo

  • Process Node: TSMC N5/N5P HPC

  • Cores: 128 Zen \4c Cores

  • Software-compatible with Zen 4

  • Optimised for scale-out performance

  • Same socket and platform as Genoa

  • Significantly Improved Energy Efficiency

  • Density-optimised Cache Hierarchy

  • On Track for 2023H1 Production & Launch

References