📅 Term: Every Fall
⏰ Time: TT 11:30 AM - 12:45 PM
Course Description
The next-generation hybrid cloud platform is integrated with Network Functions (NFs) across diverse network environments, ushering in a new computing paradigm known as in-network computing (INC). This introductory course offers system and network engineers the opportunity to engage in large-scale system optimization - learning practical data plane programming skills, gaining hands-on experience in hardware programming using software techniques, and exploring real-world applications through case studies in data centers, distributed AI/ML systems, security, and more.
Why In-Network Computing?
New Trends & Opportunities
Google, Amazon, Microsoft, NVIDIA, and others are improving large-scale AI systems through network programmability - a rapidly growing field critical in the job market.
Practical Skills
Learn the powerful P4 language (C++ like) to customize hardware functions of commercial NICs and switches for specific needs.
Hands-On Experience
Progressive exercises with a virtual machine pre-equipped with all necessary software for immediate data plane programming.
Case Studies
Deploy P4 programs for data center load balancing, network monitoring, security, and more through real-world examples and critical analysis.
Topics
- Software-Defined Networking (SDN)
- Network Function Virtualization (NFV)
- Concept of In-network Computing
- Network programmability with P4
- Approximation algorithms for P4
- INC for security (botnet, DDoS, intrusions)
- INC for data centers (load balancing, caching)
- INC for distributed systems (IoT, smart grid)
- INC for ML and AI (DT, RF, RNN)
Industry Trends & Why INC Matters Now
In-Network Computing is no longer a research curiosity - it is actively shaping how the world's largest tech companies build AI infrastructure. Here are the latest developments:
NVIDIA BlueField-4 DPU — GTC 2025
NVIDIA unveiled BlueField-4, delivering 800 Gb/s throughput and 6x the compute of its predecessor. It handles networking, storage, security, and AI orchestration entirely off the host CPU. Nearly 50% of cloud providers now rely on DPUs - Oracle, CoreWeave, xAI, and Lambda are already building on it.
Meta Open Programmable Fabric for AI — Oct. 2024
Meta revealed its Disaggregated Scheduled Fabric (DSF) at OCP Summit 2024 - a vendor-agnostic, open programmable network for AI training clusters. By 2025, it scaled to a 2-stage architecture supporting 18,432 XPUs, powered by a 51-Terabit Minipack3N switch for their largest AI clusters.
Microsoft Azure Boost SmartNIC — Deployed 2024
Microsoft's Azure Boost (FPGA-based SmartNIC) offloads all networking and storage virtualization from the host CPU - achieving 12.5 GB/s throughput, 650K IOPS, and sub-15 microsecond VM-to-VM latency. Deployed across 1 million+ Azure servers with all future VMs planned to use it.
Google Jupiter: 13 Petabits/sec Datacenter Fabric — 2024
Google's Jupiter network fabric now scales to 13 Pb/s of bisection bandwidth - a 5x capacity improvement - using software-defined networking and Optical Circuit Switches (OCS) to dynamically reconfigure topology for AI/ML workloads. This backbone powers Google's TPU v4 AI supercomputer clusters.
