[LogicFailsMe]

They're Turing complete. What else do you need?

[Our_Benefactors]

It’s possible (likely, even) to have a chip fast enough for inference, but not fast enough or with enough memory to do meaningful training runs. Like the current DGX spark.

[Npovview]

There is a reason why Google has tpu8i and tpu8t

[porphyra]

technically in order for something to be turing complete it needs infinite memory

[LogicFailsMe]

The more (memory) you buy, the more you save!