PDF 原檔：報告_OCP2026_AMD_HPE_開放標準AI機櫃_20261015_original.pdf

講者公開簡報（OCP 官方 Recorded Content 頁提供下載）。文字以 PyMuPDF 逐頁抽取，### p<n> 為頁碼；圖表內容見原檔。

圖片清單（已驗證 2026-10-19）

原始內容

p1

Mike Wong Senior Director, Product Management HPE Shangxin Du Director Product Management AMD HPE and AMD: BUILDING AI RACK SYSTEM WITH OPEN STANDARDS AI OPEN DATA CENTER

p2

AMD Helios AI Rack by HPE – Open Rack scale AI Infrastructure 72 GPUs Per rack 31 TB HBM4 Memory 260 TB/s Scale Up Bandwith 2.9 Exaflops AI Compute 1.7 PB/s HBM4 Bandwidth 43 TB/s Scale Out Bandwidth

p3

Helios Scale-up Pod Architecture EPYC MI455X … 800GbE NIC 800GbE NIC 8 16 MI455X … … … . . . 72 GPUs 18 Compute Trays . . . Scale Up Switch UAL 8 @ 128G 72 UALoE @ 212G Scale-Out Ethernet Network Up to 2.4 Tb/s/dir per GPU … … Scale Up Switch Ethernet 4 @ 200G 12 TH6 switches (512-lane Switch) 800GbE NIC 800GbE NIC 800GbE NIC 800GbE NIC 800GbE NIC 800GbE NIC 800GbE NIC MI455X 800GbE NIC 800GbE NIC 800GbE NIC MI455X MI455X MI455X EPYC MI455X … 800GbE NIC 800GbE NIC 8 16 MI455X … … … 800GbE NIC 800GbE NIC 800GbE NIC 800GbE NIC 800GbE NIC 800GbE NIC 800GbE NIC MI455X 800GbE NIC 800GbE NIC 800GbE NIC MI455X MI455X MI455X Scale-up UALoE Network 1.8 TB/s/dir per GPU Shared Load/Store Access Across Pod

p4

Compute Tray Compute Tray Scale-up Switch Tray gRPC Agents Node Manager Switch Manager Topology Service Cluster Management Daemon Event Hander Config Manager AFM API Server • Streamlined workflow for scale-up network provisioning • Fabric-level perspective for effective control, monitoring, and troubleshooting • Simplified provisioning with zero-touch scale-up network deployment • Faster implementation • Reduced need for specialized skills • Consistent adherence to best practices • Support for both advanced and turnkey deployment scenarios Switch-tray Helios Scale-up Fabric AMD Fabric Manager, Rack-wide Intelligence

p5

Operational Remediation – Link Level Retry UALoE Port Switch UALoE Frame ACK UALoE Frame NACK Retry Frame Buffer Checksum • Frame Transmission with Sequence Numbers • Each frame exchanged between nodes is assigned a local sequence number, which remains hidden from the upper protocol layers. • Frame Buffering • The sender stores transmitted frames in a local retry queue, maintaining them in a buffer until confirmation of receipt is received from the receiver. • Acknowledgment • Upon successful reception of each frame, the receiver sends an acknowledgment (ACK) back to the sender. • Loss Detection • If a frame is lost or corrupted, the receiver detects a discontinuity in the sequence numbers and responds with a Negative Acknowledgment (NACK). • Additionally, if the sender does not receive an ACK or NACK within a specific time interval, it triggers a timeout mechanism to retry transmission.

p6

Operational Remediation – vPod Compute Tray # 0 GPU

0

GPU

1

GPU

2

GPU

3

Compute Tray # 1 GPU

0

GPU

1

GPU

2

GPU

3

Compute Tray # 2 GPU

0

GPU

1

GPU

2

GPU

3

Compute Tray # 3 GPU

0

GPU

1

GPU

2

GPU

3

Compute Tray # 4 GPU

0

GPU

1

GPU

2

GPU

3

Compute Tray # 5 GPU

0

GPU

1

GPU

2

GPU

3

Job #11 Job #24 Job #13 Job #27 • A vPod can be created by combining any arbitrary set of GPUs. • Configuring a vPod necessitates DF reconfiguration. • At FCS, allocation is restricted to the granularity of the compute tray. • Communication is confined within the vPod. • Traffic is segregated in scale-up switches through VLANs. • A link failure within one vPod does not affect GPUs in other vPods. • Job assignment to a vPod: • Each vPod serves as an isolated failure domain. • Failure recovery within a vPod affects all jobs running on it. • Assign a vPod to a job when it requires equal or more than 4 GPUs.

p7

AINIC Programmable backend and frontend AI NIC is designed to run programmable RDMA over Ethernet and network offloads to support low latency, high bandwidth storage and collective communications Examples: • Software-Defined Networks (SDN) • RDMA (RoCEv2, MRC, MetaRoCE, UEC) • Security • Storage (NVMeoF) Deployed in multiple modes • Front-End Ethernet NIC • Scale-out RDMA NIC • Storage Initiator/ Target NVMeoF Vulcano 800 AI NIC Compute Tray Compute Tray Compute Tray Switch Tray Helios Rack AI Solution

p8

Scale-out Transport is evolving Network transport is evolving to adapt to the requirements RoCEv2 Reliable RDMA delivery Window Based CC(MRC) Lossy-fabric tolerance Multipath, SACK(MRC, UEC) Transport Efficiency Multiplane, Path Probing(MRC) Failure resiliency K/V cache transfer + Token Dispatch Traffic Convergence AMD Pensando P4DMA Engine One programmable data path carries every layer configurable packet & memory parsing custom logic for new protocols custom collectives in the data path telemetry that adapts as transports change

p9

MRC delivering key performance enhancements Multipathing & In-Order-Delivery 1 n Packet Spray Fast Job Completion Time • Eliminating Hash Collisions • Maximize fabric utilization with 1:1 subscription • Microseconds failure recovery • Reduce the tail latency Selective Retransmission CWND Selective Ack X X Pkt Tx Operation Efficiency • Massive bandwidth conservation • Faster lost recovery Path Aware NSCC CWND ECN Pkt TX Open ecosystem • Avoid the head-of-line blocking • Path-aware throttling • Fast ramping for bursty workload

p10

Multi-Plane Improves Redundancy & GPU Utilization MI455X Vulcano 0 1x 800G Vulcano 1 1x 800G Plane 0 1.6T OSFP 1.6T 8x8 Shuffler  Port breakout to split 800G NIC into 4×200G, distributed over four independent fabrics/planes  Packet spray & in-order delivery: achieve non-blocking GPU bandwidth for large clusters without a core layer  Highly available with better fault isolation: Maximize GPU cluster scale at lower cost  Fabric shuffler reduces the connectivity complexity Plane 1 1.6T Plane 2 1.6T Plane 3 1.6T Plane 4 1.6T Plane 5 1.6T Plane 6 1.6T Plane 7 1.6T

p11

• Massive growth in AI demand drives the need for an open rack-scale AI systems • AMD Helios rack co-designed with emerging AI workloads • 72 GPUs operate as a unified compute engine • ORW rack provides integrated power, cooling, and management • Scale-up network: • Shared memory ld/st access to over 30TB of HBM4 at 1.8 TB/s/dir per GPU • UALoE protocol provides flexible partitioning, security isolation, and multi-layer resiliency • Scale-out network: • Pensando AI-NICs provide protocol flexibility, advanced load balancing & congestion control, and telemetry & error management • 64x 1.6T Ethernet switches support large radix, multi-plane, standard-based load balancing Summary

p12

Thank You!