Table of Contents

Enum GpuDirectMessagingMode

Namespace
Orleans.GpuBridge.Abstractions.K2K
Assembly
Orleans.GpuBridge.Abstractions.dll

GPU direct messaging mode for K2K communication.

public enum GpuDirectMessagingMode

Fields

CpuRouted = 0

All messages route through CPU (default, always works). Latency: 2-10μs per message.

GpuDirectRdma = 5

Use GPUDirect RDMA for cross-node communication. Requires InfiniBand or RoCE network adapters.

InfinityFabric = 4

Use AMD Infinity Fabric for AMD GPUs. Latency: 100-300ns per message.

Use NVLink for NVIDIA GPUs with NVLink connectivity. Latency: 100-200ns per message.

PciExpressP2P = 2

Use P2P over PCIe when devices support it. Latency: 500ns-1μs per message.

PreferP2P = 1

Prefer P2P when available, fall back to CPU routing. Automatically detects and uses the best available path.