NVIDIA H100 AI accelerators based on the Hopper microarchitecture, used in the artificial intelligence and high-performance computing segments, are not enough for everyone due to the huge level of demand. Meanwhile, NVIDIA is already preparing customers for the fact that the next generation of Blackwell products will also be in short supply.
“We expect supply of our next-generation products to be limited as demand outstrips supply,” NVIDIA Chief Financial Officer Colette Kress said in a call with investors. The next-generation NVIDIA B100 accelerators based on Blackwell architecture are expected to significantly outperform current Hopper-based solutions in AI-related tasks in terms of performance and efficiency. Given the huge demand for high-performance AI accelerators in the market, it is likely that existing NVIDIA customers have already placed pre-orders for Blackwell series products.
The question is how quickly the company can increase production of SXM modules and B100 PCIe cards, as well as DGX servers based on them – these are completely new products that use completely new components. According to unconfirmed information, Blackwell will be the first NVIDIA architecture with a chiplet layout. On the one hand, this will help simplify the production of Blackwell graphics at the silicon level, since it is easier to maximize the yield of smaller chips; on the other hand, chiplet components are more difficult to package.
In addition to the flagship B100, NVIDIA is preparing a special version of the B40 for corporate clients and educational institutions, the GB200 model with the B100 graphics processor and the Grace Arm processor, as well as the GB200 NVL for training large language models. In November, NVIDIA introduced the H200 accelerator – based on the Hopper architecture, but with increased bandwidth and memory – and is now increasing its production. But the company cannot yet fully satisfy the demand for the new product.
If you notice an error, select it with the mouse and press CTRL+ENTER.