Nvidia rolls SuperNIC 1.6T per GPU; BlueField 4 will double the throughput of its predecessor and offer six times the compute power.
Networking technology took center stage at Nvidia’s GTC DC developers show in Washington, D.C. as a range of networking products were introduced for high-performance AI inference processing and more.
The Nvidia networking spotlight was on the launch of the ConnectX-9 SuperNIC, part of the Vera Rubin platform. This next-generation network interface card delivers 1.6T GPU, with advanced RDMA capabilities and PCIe Gen 6 support.
[ Related: More Nvidia news and insights ]
The SuperNIC is specifically designed for AI workloads, from training trillion-parameter models to running massive inference jobs. Security is built in at line speed, with hardware crypto acceleration and advanced features for programmable I/O, letting developers tailor infrastructure for future workloads rather than one size fits all security, according to Nvidia.
Nvidia also introduced BlueField 4, a next-generation processor that acts as the operating system for AI factories. It delivers 800Gbit/sec of throughput, double the throughput of its predecessor BlueField 3, and six times more compute than BlueField 3.
BlueField 4 combines Arm-based CPUs with the ConnectX-9 SuperNIC to accelerate storage, networking, and security across AI infrastructure. In a briefing with journalists on Monday, Dion Harris, a senior director of AI platforms at Nvidia, said BlueField 4 is “designed to power the operating system of AI factories” and will deliver six times more compute compared to its predecessor BlueField 3.
That’s because BlueField 4 received a replacement CPU: Nvidia’s own Grace CPU, a 64-core design based on the Neoverse enterprise architecture. By contrast, Bluefield 3 was based on a 16-core architecture most used in smartphones.
BlueField 4 is designed to offload and accelerate networking, storage and security workloads from the server’s host CPU and will debut in Nvidia’s Vera Rubin rack-scale servers due next year. It will be available both for Nvidia platforms like DGX and NVL 72 as well as third party server providers.
Harris said ConnectX-9 with its advanced RDMA and 48 switch lanes with PCIe Gen 6 connectivity combined with BlueField 4 “means we’re unlocking peak efficiency for AI workloads, from training trillion-parameter models to running massive inference jobs.”
More Nvidia news:
- Quantum Circuits brings dual-rail qubits to Nvidia’s CUDA-Q development platform
- Should enterprise developers care about Nvidia?
- Nvidia, Infineon partner for AI data center power overhaul
- Nvidia’s DGX Spark desktop supercomputer is on sale now, but hard to find
- Inside Nvidia’s ‘grid-to-chip’ vision: How Vera Rubin and Spectrum-XGS advanceAI giga-factories
- Nvidia and Fujitsu team for vertical industry AI projects
- Nvidia and OpenAI open $100B, 10 GW data center alliance
