NVIDIA has long been a leader in the artificial intelligence chip market. Its accelerators are the best in their class, but the most significant advantage is the CUDA software interface, writes The Motley Fool.
For many years, developers have been creating CUDA-based frameworks for the NVIDIA architecture, which, in fact, has tied many data center providers to using chips from this particular manufacturer. But Alphabet, Meta✴ Platforms and Oracle could shake NVIDIA’s strong competitive advantage with a number of new developments.
Last month, Alphabet introduced a technology stack TorchTPU open source that allows you to run PyTorch workloads on native tensor processors (TPU), while requiring only minimal changes to the code. Amazon also announced TorchNeuron’s own software environment, providing support for PyTorch code on its chips Trainium. Meta✴ established cooperation with both companies in order to be able to use alternative accelerators and not depend solely on NVIDIA products.
It should be noted that just because PyTorch can now easily run on Alphabet TPUs or Trainium chips does not mean that the performance will be as good as on NVIDIA GPUs, since PyTorch is optimized to run on CUDA and NVIDIA hardware. However, recent advances open the way for further optimizations and easier code portability, and competing chips could become much more powerful in the future. And, for example, Chinese companies have no choice at all, so they either adapt chips for CUDA, or entirely refuse from the NVIDIA software ecosystem.
Image Source: NVIDIA
Earlier this year, Alphabet started selling TPU-based systems to third-party companies, and recently introduced accelerators TPU 8tx for corporate use. In addition, from 2027 TPU will become available in the cloud Oracle Cloud Infrastructure (OCI). This will allow Oracle customers to offload some of their workloads to TPUs – which could provide better performance, better pricing (or both) compared to NVIDIA GPUs. Moreover, this gives a chance to take advantage of TPU for those developers who are “tied” to Oracle due to hosting their databases in OCI. Amazon now too ready to sell Trainium third party data center.
The opportunity for providers to get a high return on investment by using alternative accelerators for AI tasks further weakens NVIDIA’s position. The situation for NVIDIA is further complicated by the fact that thanks to the development of the PyTorch platform, specialized chips – for example, from Alphabet or Amazon – are becoming more efficient and in demand. NVIDIA’s dominance in the data center hardware segment is waning, and recent events indicate that its lead over competitors may not be as significant as some investors believed, summed up Fool.com.
If you notice an error, select it with the mouse and press CTRL+ENTER. | Can you write better? We always welcome new authors.
Source:
