The integration of NVIDIA Spectrum-X Ethernet networking technologies for AI into server lineups by Dell Technologies, HPE, and Lenovo will be the first, enabling business clients to expedite generative AI workloads. 1.6 times better networking performance for AI communication than conventional Ethernet solutions is what NVIDIA Spectrum-X promises - a new type of Ethernet networking designed specifically for generative AI.
Combining Spectrum-X with NVIDIA Tensor Core GPUs, NVIDIA AI Workbench software, and NVIDIA AI Enterprise software, the new systems from three of the leading system manufacturers would provide businesses the building blocks they need to use generative AI to revolutionize their businesses.
According to NVIDIA founder and CEO Jensen Huang, "Generative AI and accelerated computing are driving a generational transition as enterprises upgrade their data centers to serve these workloads. To accelerate the transition to the era of generative AI, NVIDIA's top server manufacturer partners are releasing a new wave of systems, and accelerated networking is the catalyst for them."
AI Acceleration Through Networking Purpose-Built
The NVIDIA BlueField-3 SuperNIC, a new class of network accelerators for supercharging hyperscale AI workloads, together with acceleration software, are combined with the “exceptional performance” of the Spectrum-4 Ethernet switch to achieve peak efficiency for AI workloads. One of the most sophisticated infrastructure computing platforms in the world, BlueField-3 DPUs, would be enhanced by NVIDIA Spectrum-X.
The 51Tb/sec Ethernet switch designed specifically for artificial intelligence, Spectrum-4 would minimize network congestion for multi-tenant AI cloud applications while offering very efficient data throughput both under load and at scale. Maximum network infrastructure utilization is always possible thanks to its intelligent, well-tuned routing technology.
To improve performance for AI training and inference traffic on the east-west network inside the cluster, BlueField-3 SuperNICs are built for network-intensive, massively parallel computing. They would provide up to 400Gb/s RDMA over Converged Ethernet (RoCE) network access between GPU servers. Additionally, they would make multi-tenant, secure data center systems possible while maintaining predictable and segregated performance amongst tenant tasks. BlueField-3 SuperNICs ca be a good fit for “enterprise-grade” servers because of their half-length, half-height, and power-efficient PCIe form factor.
NVIDIA software development kits including Cumulus Linux, Pure SONiC, and NetQ - which collectively propel the platform's breakthrough performance - as well as the NVIDIA DOCA software framework, which forms the basis of BlueField, are among the acceleration software components powering NVIDIA Spectrum-X.
Frameworks, pretrained models, and development tools are all provided by NVIDIA AI Enterprise for production AI that would be secure, reliable, and maintained. Using a PC or workstation, developers can rapidly build, test, and tweak pretrained generative AI models using NVIDIA AI Workbench. These models can then be scaled to almost any data center or cloud environment.
Chairman and CEO of Dell Technologies Michael Dell said, "Building systems to meet the demands of large language models and generative AI applications requires accelerated computing and networking. Dell Technologies and NVIDIA are collaborating to offer customers the software and infrastructure required to swiftly and safely extract intelligence from their data."
According to Antonio Neri, President and CEO of HPE, "generative AI will undoubtedly drive innovation across multiple industries. To accommodate a range of dynamic workloads, these potent new applications will need a radically altered architecture. HPE and NVIDIA are working together to provide systems with the necessary strength, efficiency, and scalability to support these applications, allowing clients to fully grasp the promise of generative AI.”
"Even though it puts unprecedented demands on enterprise infrastructure, generative AI can power unprecedented transformation," said Lenovo Chairman and CEO Yuanqing Yang. “Lenovo and NVIDIA are collaborating together to develop systems that are fast, efficient, and equipped with the networking, processing, and software required to run contemporary AI applications.”
Supercomputer NVIDIA Israel-1
Utilizing Spectrum-X as its power source, the NVIDIA Israel-1 supercomputer serves as a model architecture for artificial intelligence systems of the future. Israel-1, a partnership with Dell Technologies, makes use of BlueField-3 DPUs, SuperNICs with Spectrum-4 switches, and Dell PowerEdge XE9680 servers driven by the NVIDIA HGX H100 eight-GPU arihitecture.
Availability
The first quarter of 2024 is likely to see the release of new systems from Dell, HPE, and Lenovo that include the whole NVIDIA AI stack.
