3 ms·
Nvidia will face more competition in the future, but it has technological edge. So far Nvidia has build microarchitectures that are generic. They just put ten
by Nokinside 9y ago
Nvidia will face more competition in the future, but it has technological edge.
So far Nvidia has build microarchitectures that are generic. They just put tensor cores, F64, F32, F16 units and 8-bit inference support into the same GPU. They are used in HPC, graphics, DL inference and training. That's four different domains.
I suspect that in the future Nvidia builds 2-3 different microarchitectures.
One for HPC + graphics.
One for DL training.
One for DL inference.
- Hydraulix989 9y agoRight, I wonder just how specialized these chips are. There are still changes in modeling and architecture coming hot off the research presses. It's easy-ish to hard code a specific architecture in your Verilog and claim massive efficiency improvements. While AI is still very much in the research stage, general-purpose is what really matters.