NVIDIA announced H100 systems will ship bundled with proprietary AI software, creating an integrated hardware-software product offering.
Nvidia plans to bundle a five-year license of its commercial AI Enterprise software with every PCIe-based H100 GPU coming to servers soon, a strategy disclosed at the Supercomputing 2022 conference ahead of the launch of its next-generation H100 datacenter GPU. The company has promised the H100 will provide substantial performance improvements over the A100. This bundling initiative is part of Nvidia's broader ambition to build a multibillion-dollar software business complementing its silicon revenue.
The H100 will be available in servers from numerous manufacturers. Dell will offer new PowerEdge systems including the XE9680, which uses Nvidia's HGX board to pack eight H100 or A100 GPUs along with two 4th-gen Intel Xeon Scalable processors in an air-cooled design. Supermicro is releasing its "most advanced GPU server yet" with its new 8U Universal GPU server equipped with eight H100s, while HPE's Cray supercomputer unit will support the H100 with the XD6500. Other vendors planning H100-based servers include Asus, Atos, Gigabyte, Lenovo, Penguin Computing, QCT, and Ingrasys, with first-quarter 2023 availability expected for some.
Nvidia AI Enterprise is a bundle of AI tools including frameworks like PyTorch and TensorFlow along with offerings such as Nvidia Inference Server, optimized to run in containers or virtual machines on VMware's vSphere or Red Hat's OpenShift platforms. A five-year subscription license normally costs $8,000 per CPU socket. By bundling this license with H100 PCIe GPUs, a dual-socket server with two H100 PCIe cards would be fully covered for Nvidia AI Enterprise licensing, though servers with more CPU sockets than H100 PCIe cards would require additional purchases.
Nvidia offers multiple licensing paths. Subscription licenses include standard support services, with critical support starting at $450 for one year and reaching $2,250 for five years. Perpetual licenses are available with one, three, and five-year support plans ranging from $4,494 to $80,90 per CPU socket. Nvidia is bundling the software specifically with the PCIe version of the H100 because this form factor fits mainstream servers, unlike the SXM form factor used in Nvidia's DGX systems and HGX motherboards.
According to Dion Harris, head of datacenter product marketing at Nvidia: "By including that with our H100 platform, it's basically offered as a value-add, and we want to encourage customers to deploy and use this in production. And so this will allow us to give them the certainty that we'll be standing behind them." The strategy depends on businesses recognizing value in the commercial software, with the expectation that free users will eventually pay for expanded deployments.