• Alibaba Cloud and Cambricon joined the PyTorch Foundation as Platinum members, while Ant Group joined at Gold level
  • Platinum status gives Alibaba Cloud and Cambricon seats on the Governing Board and Technical Advisory Council

The fact

Alibaba Cloud and Chinese AI chipmaker Cambricon joined the PyTorch Foundation as Platinum members on 8 September. Ant Group joined as a Gold member. The announcement was made at PyTorch Conference China in Shanghai.

Platinum membership gives Alibaba Cloud and Cambricon a seat each on the foundation’s Governing Board and Technical Advisory Council. Both companies are expected to contribute to work on AI infrastructure and accelerator support. Huawei also attended the conference as an existing PyTorch member and contributor.

PyTorch is used to build and run AI models on different types of computing hardware. For an accelerator to work with PyTorch, the supporting software must be able to recognise the device and run the operations required by a model.

Running a model across several devices is more complicated. Those devices also need software that allows them to communicate and work together. These are general PyTorch requirements and do not suggest that Alibaba Cloud or Cambricon has a specific compatibility problem.

The assessment

Membership gives Alibaba Cloud and Cambricon a formal place in PyTorch governance, but it does not tell a customer how well a model will run on a particular system. That still depends on the software connecting PyTorch to the accelerator and the wider computing environment.

The operations used by a model need to be supported on the accelerator. Where an operation cannot run natively, software may use another execution path, including the CPU, which can keep the model working while changing performance. Larger training jobs add another requirement: communication software must coordinate the workload across multiple devices. A successful test on one accelerator therefore does not establish how the same workload will perform across a cluster.

For BTW readers, the membership matters only if it leads to better support for real workloads. That means more operations running natively on the accelerators, stable distributed training across multiple devices and software that continues to work as PyTorch evolves. The board seats give Alibaba Cloud and Cambricon a voice in the project; customers will judge the result by what they can run reliably on their systems.

What to watch

Watch for new PyTorch integrations, operator support and distributed-training releases from Alibaba Cloud and Cambricon. Reproducible workload tests should show which operations run natively, whether fallback paths are used and how jobs perform across multiple accelerators. Those results would provide a clearer measure of practical support than membership status alone.