Future Challenges for HPC Networking @Scale
summary
Networks shift toward commodity Ethernet with Slingshot (Ethernet) deployed on Exascale machines and the Ultra Ethernet Consortium forming, but MPI still dominates MPI 98% and RDMA/IB verbs and portals remain critical. Frontier runs ~10,000‑node Frontier configurations with CPUs, GPUs, multiple memories and on‑package links, creating integration pain: GPUs lack consistent NIC integration, link rates approach one packet per ~1.2 ns, and smart NICs wind up used as smart NICs for app offload rather than network abstractions. Key challenges are latency and small‑message rate, NIC parallelism, resiliency, congestion control and richer active‑message support.
AI-written summary · about
chapters (9)
- 0:00HPC networking challenges
- 1:50Hardware evolution in HPC
- 5:19Modern HPC systems
- 9:13SmartNICs limitations
- 10:59HPC application landscape
- 13:25Software trends in HPC
- 16:53GPU and network integration
- 17:58Ultra Ethernet future
- 18:59Future HPC networking challenges
related talks
31mLive Panel: Scale-Out Networking for AI Supercomputers
19mAI/ML Networking Challenges: The Fast and the Finicky!
47mThe World of 100G Networking
26mApplying New High-Throughput Network Solutions in IEC (Integrated Edge Cloud)
13mLive Q&A Session #2 – Shashi Gandham
25mKeynote: Network Transformation: Are We Ready for 5G and Edge Clouds?
24mScalable Small Message Aggregation on Modern Interconnects
12mPreview: Networking