Composable AI Solution For Power 9
summary
H3 Platform presents Falcon II a PCIe Gen4 fabric GPU expansion box that connects to Power9 servers and loads up to 16 GPGPUs shared across up to four hosts. Demo shows live composable allocation — composable allocation 16:1 8+8 4x4 — hot‑plugging and peer‑to‑peer GPU transfers that stay off host memory, with a management GUI exposing topology and per‑slot telemetry. Known limits include PCIe bus and firmware constraints, driver support for hot‑add/remove, and practical scalability hitting about PCIe Gen4 fabric ceilings around sixteen to thirty‑two GPUs. System marketed as Falcon II.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:40Requirements
- 2:18GPU to Server
- 12:04Interchange
- 13:13Limitations
- 15:56GPU
- 26:07Dynamic Cloning
- 28:44Demo Environment
- 35:14GPU Placement
- 40:02Monitor PCIe
related talks
14mServing Frontier MoE Models at the Lowest Token Cost
38mDynamic Scaling of GPUs for Container Apps with Composable Disaggregated...
13mDecentralized ZettaFlop Uber GPU/XPU Network for Al and Spatial Computing
30mBeyond the DC Walls: Building a Nationwide GPU Fabric With KubeVirt and Wide-Area…
04mLightning Talk: Dynamic GPU Fraction and Sharing With Cloud Native Principle
32mA Deep Dive on Supporting Multi-Instance GPUs in Containers and Kubernetes
23m1.2 Tb/S Packet Processing of Radio Telescope Data on a GPU
08mContainerization – It’s Not Just for CPUs Anymore: Introducin...