If compute is not the bottleneck, memory is easy-ish to produce (the hard part is mostly on the fab side); what stops a Chinese NVIDIA (huawei) from being 10x cheaper?
Making memory is easy. Packaging that memory within a few millimeters of a piece of silicon using TSVs and maintaining signal integrity on a 1024 bit bus is really, really hard.
LLMs aren’t all that compute constrained or even memory constrained. It’s just that pushing dozens of terabits per second through a piece of silicon is a physics problem.
Comments
If compute is not the bottleneck, memory is easy-ish to produce (the hard part is mostly on the fab side); what stops a Chinese NVIDIA (huawei) from being 10x cheaper?
Making memory is easy. Packaging that memory within a few millimeters of a piece of silicon using TSVs and maintaining signal integrity on a 1024 bit bus is really, really hard.
LLMs aren’t all that compute constrained or even memory constrained. It’s just that pushing dozens of terabits per second through a piece of silicon is a physics problem.
I think it's mostly the ramp-up time, but ChangXin Memory Technologies (CXMT) is basically aspiring to do just this.