NVIDIA has revealed the latest business collaboration in the form of the new Spectrum-X Ethernet networking technologies designed to speed up generative AI workloads even more.
Joined by Dell Technologies, HP Enterprise, and Lenovo, this new class of Ethernet networking is capable of boosting networking performance for AI communication by up to 60% and will be offered by the aforementioned system makers who already have their own platform offering NVIDIA’s Tensor Core GPU-based systems with NVIDIA AI Enterprise and AI Workbench software for an even more holistic generative AI ecosystem.
In terms of hardware, the Spectrum-4 Ethernet switch and the in-house BlueField-3 SuperNIC which is a brand new class of network accelerators working together with dedicated software to deliver some of the most advanced infrastructure to date.
For reference, the Spectrum-4 is capable of doing 51Tbps Ethernet for superior data throughput at scale and under load while minimizing congestion for multi-tenant environments. Crafted with several dynamically adjusted software, intelligent fine-tuning will make sure the system operates at maximum potential at all times.
Meanwhile, the BlueField-3 SuperNICs offer up to 00Gb/s RDMA over Converged Ethernet (RoCE) network connectivity between GPU servers while boosting performance for AI training and inferencing that is both secure and efficient.
Lastly, the innate acceleration software within Spectrum-X is ready to support dev kits including Cumulus Linux, Pure SONiC, and NetQ.
Availability
Spectrum-X-enabled systems offered by Dell Technologies, HP Enterprise, and Lenovo are expected to be released during Q1 2024.
Just a quick note, the NVIDIA Israel-1 supercomputer which is a project developed in collaboration with Dell Technologies also utilizes Spectrum-X with the baseline hardware powered by Dell PowerEdge XE9680 servers containing 8x NVIDIA HGX H100 GPU + BlueField-3 DPUs and SuperNICs with Spectrum-4 switches.
