The IBM AC922 system consists of 4,608 compute servers, each containing two 22-core IBM Power9 processors and six NVIDIA Tesla V100 graphics processing unit accelerators, interconnected with dual-rail Mellanox EDR 100Gb/s InfiniBand. Summit also possesses more than 10 petabytes of memory paired with fast, high-bandwidth pathways for efficient data movement. The combination of cutting-edge hardware and robust data subsystems marks an evolution of the hybrid CPU–GPU architecture successfully pioneered by the 27-petaflops Titan in 2012.
This is technically an area I work in, so I can provide some commentary. This thing is very similar (except for using IBM power9) to what is used to train self driving cars.
Ironically that you say that, but there is so much data that comes out of a car, that they remove the drives and ship them to the data center because it costs to much to use the internet when you have petabytes of data to move.
Kind of. A minicomputer is not going to cut it. (and this isn't a mainframe) This is a cluster of somewhere around 100k CPUs.
Not exactly a practical thing. You'd have to have maybe 15-20 semi trucks running in a row following behind the car, and a few more with generators and fual.
*fuel
Not so much internet as the architecture speed, storage density and overall size.
The architecture everyone is using is really only a generation past what they used 2 decades ago. It is only just now exploding in the past 5 years now that GPUs are fast enough to get results relatively quickly.
Traditional HPC clusters really can't be re-purposed for AI (lack of GPUs), so there is a huge investment needed for anyone entering the field.
except that is a chicken-egg problem.
Nothing is really 100%
There is always the possibility of an unknown unknown, like a meteorite impact or something.