KV Cache Transfer Over RDMA Between Prefill and Decode Nodes
Push mode RDMA writes and pull mode reads create a 98x throughput gap on identical hardware.
Tobias Wrenfeld
Section
High-Speed Interconnects
4 stories in High-Speed Interconnects.
RoCEv2 Congestion Control for AI Storage Traffic
Configuring ECN, DCQCN, and PFC settings correctly is essential for preventing fabric bottlenecks.
Nikolai Vereshchagin
How RDMA Bypasses the CPU to Sustain All-Reduce Throughput in GPU Cluster Training
Bypassing the CPU with RDMA and GPUDirect lets networks, not software, limit training speed.
Tobias Wrenfeld
RDMA Fundamentals for LLM Infrastructure Engineers
RDMA removes CPU bottlenecks from LLM data movement across inference and training.
Priya Subramaniam