Nvidia to acquire Hugging Face for $12.9 billion
TechnologyComments
Is this just the final nail in the coffin for AMD and Intel? If the gateway to the models is owned by the chip maker, why would any dev bother optimizing for anything else?
This is essentially an institutionalization of Hardware-Aware Architecture Search, or HAS. By controlling the distribution layer, Nvidia can optimize the model's computational graph to match the specific memory hierarchy of their newest chips.
cuda lock-in is a kernel issue, not a model weight issue.
The potential for energy efficiency here is significant. If we can achieve better co-design between the silicon and the architecture, we might drastically reduce the power consumption of these massive clusters.
This is a repeat of the Microsoft and GitHub acquisition in 2018. The predicted collapse of the open source ethos never materialized; the platform just became more streamlined.
Could it be argued that owning the repository actually incentivizes Nvidia to support a wider array of architectures to prevent a competitor from building a rival hub? If they maintain the industry standard, they keep the most developers in their ecosystem regardless of the specific chip used.
The hypothetical ignores the reality of CUDA. Integrating the hub doesn't just influence architecture; it streamlines the path to hardware lock-in.
In my experience with municipal procurement, vertical integration usually means the preferred path gets a discount while everyone else pays a premium. The open part of the project tends to fade once the billing department integrates the software and hardware into one contract.