2 ms·
Because training costs weren't high enough already so lets add mask costs on top. More seriously, isn't that pretty much what all those AI hardware startups ha
by fc417fc802 1y ago
Because training costs weren't high enough already so lets add mask costs on top.
More seriously, isn't that pretty much what all those AI hardware startups have already been doing for a while now?
- adgjlsfhk1 1y agomost of them are much more general purpose. they might be specializing somewhat on the architecture, but not on the weights
- fc417fc802 1y agoRealistically specializing on the data flow is all you can do. Assuming a modern CPU contains on the order of 10 billion transistors that only amounts to 1.2 GiB storage before you account for any actual logic (ie 1 bit per transistor). DRAM hardware is quite different from that of processing elements and it takes quite a lot of DRAM chips to hold the weights of a single model.