Overview of Popular GPU/TPU Cluster Networking Technologies: NVLink, InfiniBand, RoCE, and DDC

This article reviews the main GPU/TPU cluster networking solutions—including NVLink, InfiniBand, RoCE Ethernet, and DDC full‑schedule fabrics—examining their latency, loss‑free transmission, congestion control, cost, scalability, and suitability for large‑scale LLM training workloads.

AI trainingDDCGPU networking

0 likes · 16 min read

Overview of Popular GPU/TPU Cluster Networking Technologies: NVLink, InfiniBand, RoCE, and DDC

Architects' Tech Alliance

Feb 29, 2024 · Industry Insights

Choosing the Right GPU Cluster Network: NVLink, InfiniBand, RoCE & DDC Explained

This article examines the key GPU/TPU cluster networking options—NVLink, InfiniBand, RoCE Ethernet, and emerging DDC full‑scheduling fabrics—detailing their latency, loss‑less transmission, congestion control, cost, power, and scalability considerations for large‑scale AI training deployments.

AI trainingDDC fabricGPU networking

0 likes · 18 min read

Choosing the Right GPU Cluster Network: NVLink, InfiniBand, RoCE & DDC Explained

Liangxu Linux

Mar 27, 2023 · Fundamentals

Can Linux Reach True Line‑Speed? Deep Dive into Kernel Packet Forwarding Bottlenecks

This article explains why Linux’s native networking stack falls short of line‑speed, analyzes the core latency sources such as DMA, memory copying, routing look‑ups and lock contention, and proposes a comprehensive set of kernel‑level redesigns—including VOQ queues, lock‑free data structures, and user‑space stacks—to achieve near‑wire‑speed forwarding.

DMALinuxVOQ

0 likes · 39 min read

Can Linux Reach True Line‑Speed? Deep Dive into Kernel Packet Forwarding Bottlenecks