Nvidia is well-known for its high-performance gaming GPUs and industry-leading AI GPUs. Nonetheless, the trillion-dollar GPU producer can also be recognized for controlling how its GPUs are used past firm partitions. For instance, it may be fairly restrictive with its AIB companions’ graphics card designs. Maybe not surprisingly, this stage of management additionally seems to increase past card companions over to AI clients, together with Microsoft. The Info experiences that there was a standoff between Microsoft and Nvidia over how Nvidia’s new Blackwell B200 GPUs have been to be put in in Microsoft’s server rooms.
Nvidia has been aggressively pursuing ever bigger items of the information heart pie, which is straight away clear should you take a look at the way it introduced the Blackwell B200 components. A number of occasions through the presentation, Jensen Huang indicated that he would not take into consideration particular person GPUs any extra — he thinks of all the NVL72 rack as a GPU. It is a moderately clear effort to realize further income from its AI choices, and that extends to influencing how clients set up their new B200 GPUs.
Beforehand, the shopper was chargeable for shopping for and constructing applicable server racks to deal with the {hardware}. Now, Nvidia is pushing clients to purchase particular person racks and even total SuperPods — all coming direct from Nvidia. Nvidia claims this may enhance GPU efficiency, and there is advantage to such discuss contemplating all of the interlinks between the assorted GPUs, servers, racks, and even SuperPods. However there’s additionally plenty of greenback payments altering fingers if you’re constructing knowledge facilities at scale.
Nvidia’s smaller clients is perhaps comfortable with the corporate’s choices, however Microsoft wasn’t. VP of Nvidia Andrew Bell reportedly requested Microsoft to purchase a server rack design particularly for its new B200 GPUs that boasted a kind issue just a few inches totally different from Microsoft’s current server racks which are actively utilized in its knowledge facilities.
Microsoft pushed again on Nvidia’s advice, revealing that the brand new server racks would stop Microsoft from simply switching between Nvidia’s AI GPUs and competing choices corresponding to AMD’s MI300X GPUs. Nvidia ultimately backed down and allowed Microsoft to design its personal customized server racks for its B200 AI GPUs, but it surely’s in all probability not the final such disagreement we’ll see between the 2 megacorps.
A dispute like this this can be a signal of how massive and helpful Nvidia has change into over the span of only a 12 months. Nvidia turned essentially the most helpful firm earlier this week (briefly), and that title will doubtless change fingers many occasions within the coming months. Server racks aren’t the one space Nvidia needs to manage, because the tech big additionally controls how a lot GPU stock will get allotted to every buyer to take care of demand, and it is utilizing its dominant place within the AI house to push its personal software program and networking programs to take care of its place as a market chief.
Nvidia is benefiting massively from the AI increase, which began when ChatGPT exploded in recognition one and a half years in the past. Over that very same timespan, Nvidia’s AI-focused GPUs have change into the GPU producer’s most in-demand and highest earnings producing merchandise, resulting in unimaginable monetary success. Inventory costs for Nvidia have soared and are at the moment over eight occasions larger than they have been firstly of 2023, and over 2.5 occasions larger than at the beginning of 2024.
Nvidia continues to make use of all of this further earnings to nice impact. Its newest AI GPU, the Blackwell B200, would be the quickest graphics processing unit on the earth for AI workloads. A “single” GPU delivers as much as a whopping 20 petaflops of compute efficiency (for FP8) and is 4 occasions sooner than its H200 predecessor. In fact it is really two massive chips for such a GPU, besides that is plenty of quantity crunching prowess. It additionally delivers 5 petaflops of dense compute for FP16/BF16, utilized in AI coaching workloads.
As of June twentieth, 2024, Nvidia inventory has dropped barely from its $140.76 excessive and closed at $126.57. We’ll need to see what occurs as soon as Blackwell B200 begins transport en masse. Within the meantime, Nvidia continues to go all-in on AI.
