News for hackers. Stuff that matters.

A comment on Huawei releases an open weight model trained on Huawei Ascend GPUs. Parent comment

Re:Huawei releases an open weight model trained on Huawei As... by seydor on Wednesday July 02, 2025 @05:15AM (#44441577)
even if their chips weren't as good, what stops a company from training a large model for a very long time in less capable hardware? Is there a way to overcome memory limitations somehow?
[ Parent | on HN ]
Re:Huawei releases an open weight model trained on Huawei As... by umgefahren on Wednesday July 02, 2025 @05:23AM (#44441630)
In general yes, you can (and do) shard the model over multiple GPUs. If you want to do that yourself look at DeepSpeed or FSDP . There is a communication overhead though and the speed at which the GPUs can communicate is key. Thats where NVLink comes in btw. So yes, it’s actually what you can and do do. However this limits your ability to iterate on the models quickly and from what I‘ve read a lot of times the foundational labs throw out their models because by the time they are done training they are already outdated.
[ Parent | on HN ]
Re:Huawei releases an open weight model trained on Huawei As... by checker659 on Wednesday July 02, 2025 @05:44AM (#44441759)
NVLink isn’t magic
[ Parent | on HN ]
Re:Huawei releases an open weight model trained on Huawei As... by rfoo on Wednesday July 02, 2025 @10:27AM (#44444146)
> the speed at which the GPUs can communicate is key

Guess what a telco equipment company is good at :p

[ Parent | on HN ]