Nvidia unveils HGX-2, a cloud computing platform for AI and high performance computing, which consists of 16 GPUs delivering 2 petaflops per second
Hoping to maintain the high ground in artificial intelligence and high-performance computing, Nvidia Corp. late Tuesday debuted …
Context & Ripple Effects
HGX-2 extends a line that began with the Tesla P100 in 2016, when Nvidia first pitched a datacenter GPU built for deep learning rather than graphics. By 2018 the pitch has moved up the stack: instead of selling one card, Nvidia is selling a 16-GPU cloud computing platform rated at 2 petaflops per second, aimed squarely at cloud providers serving both AI and high-performance computing customers.
First-order effects
- Cloud operators adopting HGX-2 buy a pre-integrated 16-GPU system rather than assembling their own multi-GPU servers, shifting integration work and margin toward Nvidia.
- Rival chip vendors now compete against a bundled platform-plus-software offering in the AI and HPC segments, not against individual GPUs.
Second-order effects
- Supercomputing centers follow the same path: NERSC's Perlmutter machine later scales the approach to 6,144 A100 Tensor Core GPUs (unveiled as the fastest supercomputer for AI workloads), showing that whole-system GPU procurement displaces traditional CPU-first designs.
- Each generation of platform raises the reference bar — the Hopper architecture and its 80B-transistor H100 reset what buyers expect from a single GPU, forcing competitors to match full-system performance claims rather than chip-level specs.
Third-order effects
- If the pattern holds, AI infrastructure consolidates around vertically integrated platforms: by 2023 Nvidia's DGX GH200 supercomputer platform announcement shows the same playbook extended to networking and robotics, with the vendor defining the entire compute stack rather than components.
The trend: Nvidia is executing a decade-long shift from selling discrete GPUs to defining complete AI infrastructure platforms, with each product generation pulling more of the datacenter stack under its control.