While a great deal of NVIDIA’s success in servers over the past decade has after all come from their proficient GPUs, as a enterprise NVIDIA lately is rather more than a fabless GPU designer. With extra software program engineers than {hardware} engineers on employees, it’s software program and ecosystem performs which have actually cemented NVIDIA’s place as the highest GPU producer, and created a bigger marketplace for their GPUs. At the identical time, it’s these ecosystem performs which have allowed NVIDIA to construct a profit-printing machine, diversifying past simply GPU gross sales and transferring into techniques, software program, assist, and different avenues.
To that finish, NVIDIA this morning is formally rolling out a brand new ecosystem play aimed toward high-end deep studying servers, which the corporate is branding as NVIDIA-Certified Systems. Soft-launched again within the fall, at present the corporate is giving this system a extra correct introduction, detailing this system and asserting a few of the companions. Under NVIDIA’s plan, going ahead clients can decide to purchase NVIDIA-Certified techniques if they need an additional assure on system efficiency and reliability, in addition to decide in to purchasing assist contracts to get entry to direct, full-stack technical assist from NVIDIA.
Conceptually, the certification program is quite easy, due largely to its {hardware} necessities. Systems first have to be utilizing NVIDIA’s A100 accelerators, together with Mellanox Ethernet adapters and DPUs. Or in different phrases, the servers already have to be utilizing NVIDIA silicon the place out there. OEMs can then submit techniques assembly these {hardware} necessities to NVIDIA, who will check the techniques throughout a number of metrics, together with multi-GPU and multi-node DL efficiency, community efficiency, storage efficiency, and safety (safe boot/root of belief). Systems that go these checks can then be labeled as NVIDIA-Certified.
Those licensed techniques, in flip, are eligible for added full-stack technical assist by means of NVIDIA and the OEM. Customers can decide to purchase multi-year assist contracts, which entitles them to assist by means of the OEM and NVIDIA. NVIDIA primarily assumes accountability for all software program assist above the OS, together with their {hardware} drivers, CUDA, their huge assortment of frameworks and libraries, and even main open supply libraries like TensorFlow. The latter is what makes NVIDIA’s assist proposition significantly beneficial, as they’re primarily committing to serving to clients with any type of GPU or deep learning-related software program concern.
Of course, that assist received’t come without cost: that is the place NVIDIA will probably be making their cash. While NVIDIA isn’t charging OEMs for certification (so there’s no extra certification tax baked into the {hardware}), assist contracts are priced primarily based on the variety of GPUs. In one instance, NVIDIA has acknowledged {that a} three yr assist contract for a dual-A100 system could be $4,299, or about $715 per-year per-GPU for assist. So one can think about how shortly this ratchets up for bigger Four and eight means A100 techniques, after which once more for a number of nodes.
For NVIDIA and its OEM companions, the creation of a certification program is a simple option to attempt to additional develop the marketplace for deep studying servers, particularly for mid-sized companies. The marketplace for AI {hardware} has been booming, and NVIDIA needs to maintain it that means by making it simpler for potential clients to make use of their wares. NVIDIA already has the top-end of the market lined on this respect with their direct relationships with the hyperscalers – and by extension their small-cap cloud computing clients – so a {hardware} certification program fills the center tier for organizations which are going to run their very own servers, however aren’t going to be a large buyer that…
