I find it extremely hard to believe with their level of espionage that they can’t come up with a clone of a 5 year old GPU and slam 32GB of vram on it.
The issue is access to leading edge semiconductor fab nodes and high end lithography machines (if they want to build their own leading edge node at SMIC).
Asianometry on YouTube has great content on this topic, he’s great at digesting complex industrial history
the hard part for them is actually that their fabs have a hard time making larger dies.
although china has made significant fast process on the capability of their fabs, and being able to already make 5nm yields, they have a problem with doing it with enough volume (possibly due to having high defect rate). its likely part of the reason for example, major companies like huawei havent switched fully to in country fabs for their devices.
GPU dies are absolutely huge and when you have a high defect rate on cuts of silicon, getting working dies are exponentially hard. China also does not have as many driver devs with enough experience to fully switch off it just yet. GPU drivers arent easy, and intels forray into doing dGPUs give you an example that even for multi billion dollar companies, getting a competent driver team isnt easy.
