- Joined
- Apr 24, 2020
- Messages
- 2,723 (1.60/day)
ASICs are always the solution to everything, always have been. The problem is in balancing economic benefits. An ASIC needs to be profitable if a company is to survive. Is there enough AI demand to sustain an AI-focused ASIC? Remember that a modern 3nm chip is likely a $X Billion+ investment, if not on the scale of $X0 Billions (if we include the massive software costs NVidia has incurred in making TensorFlow work with CUDA), that's a lot of zeros.
Its not like $Billion chips grow on trees. Every attempt to build a better ASIC accelerator leads to a 7nm chip (where its maybe only 100-million costs), but NVidia's 3nm $X0 Billion chip curbstomps in power-efficiency and performance anyway. How do you build a comparable chip when NVidia has a process node and money advantage?
NVidia has the benefit of relying upon the GPU market (for 64-bit scientific compute like weather modeling, physics simulations, FEA analysis, raytracing for movies, and of course video games... etc. etc.) if this AI thing ever fails. And as it turns out, a lot of AI chips have the same memory bandwidth problems that NVidia's GPUs face. So NVidia's GPUs are ideal for benefiting from AI hype, but without actually taking unnecessarily large risks like an AI-specific chip would do.
That's the thing: NVidia is the AI leader and manages to do so without taking unnecessary risks. That's an interesting position from a business perspective. I can see why people like this business. But I do think AI is bubble-hype and that this all will go down soon. So that's where my interests in NVidia lies, that the company clearly will survive an AI winter (albeit with lower valuations and lower profits).
This won't happen until after the last nodes are eek'd out. NVidia is too far ahead in raw money. When they can build 3nm when everyone else is on 7nm. And NVidia will be on 1.8nm when the rest of the world gets to 5nm. Etc. etc.
Eventually the nodes will end and ASICs will have a shot. But for now, NVidia's shear capital can sustain a lead (much like how Apple's chips sustain a process-node lead).
There's no point spending $Billion just to be beaten by NVidia anyway.
