Nvidia Corp. has announced the next generation of its Ethernet networking technology for artificial intelligence-focused servers and says Dell Technologies Inc., Hewlett Packard Enterprise Co. and Lenovo Group Ltd. will be the first to integrate it into their products.

Nvidia says its Spectrum-X technology, unveiled today, is purpose-built for AI workloads and capable of delivering 1.6 times higher networking performance than traditional Ethernet for AI communication.

The new servers announced by Dell, HPE and Lenovo combine Spectrum-X with Nvidia’s most powerful H100 Tensor Core graphics processing units, as well as its AI Enterprise and AI Workbench software. They’re said to offer a complete package for enterprises wanting to implement the most advanced generative AI models.

According to Nvidia, Spectrum-X combines the rapid performance of its newest Spectrum-4 Ethernet switch with the Nvidia BlueField-3 SuperNIC (both pictured), which is a new class of network accelerator designed for speeding up AI workloads, as well as acceleration software. The technology is said to complement Nvidia’s BlueField-3 data processing units, which are processors that offload, isolate, accelerate and secure data center services, freeing up the GPUs to focus on AI processing.

The company said Spectrum-4 is the first-ever 51-terabit-per-second Ethernet switch. It was built to ensure the highest possible data throughput at scale and under intense loads, to minimize network congestion. This is necessary because of the vast amounts of data AI workloads must process. With the addition of fine-tuned routing software, it ensures the network infrastructure is used in the most optimal way.

As for the BlueField-3 SuperNICs, these facilitate remote direct memory access speeds of up to 400 gigabits per second over a Converged Ethernet network to connect each AI server within a cluster. The company said it will have a noticeable impact on the performance of AI training and inference systems. Another benefit is high security, as the BlueField-3 SuperNICS create a secure, multi-tenant data center environment that isolates each workload.

The new systems from Dell, HPE and Lenovo will become available in the first quarter of next year, Nvidia said.

Nvidia founder and Chief Executive Jensen Huang highlighted the importance of an accelerated networking framework, saying it is the “catalyst for a new wave of AI systems from Nvidia’s leading server manufacturer partners to speed the shift to the era of generative AI.”

HPE CEO Antonio Neri explained the thinking behind the new tech, saying generative AI requires a fundamentally different network architecture to support a wide variety of dynamic workloads. To enable customers to realize AI’s potential, he said, “HPE is partnering with Nvidia to build the systems with the required power, efficiency and scalability to support these applications.”

Spectrum-X has already been installed on the Nvidia Israel-1 supercomputer, which serves as a reference architecture for companies that want to build next-generation AI systems on-premises. Israel-1 was built using Dell’s PowerEdge XE9680 servers and is powered by Nvidia’s HGX 100 platform that features eight GPUs, BlueField-3 DPUs and SuperNICs with Spectrum-4 switches.

“Through our collaboration, Dell Technologies and Nvidia are providing customers with the infrastructure and software needed to quickly and securely extract intelligence from their data,” said Dell Chairman and CEO Michael Dell.

The company’s networking acceleration software is said to enhance performance even further, providing access to software development kits such as Cumulus Linux, Pure SONiC and NetQ, as well as the Nvidia DOCA software framework that sits at the heart of its BlueField chips. AI Enterprise packages dozens of popular AI frameworks, pretrained models and development tools, while AI Workbench offers tools for customizing AI models on a PC or workstation.

Photo: Nvidia