@sanchuah if I were to hazard a guess, it might be because driving the voltage up or down on the entire bus requires non-trivial current to do so quickly.
This sort of rushing current is inefficient for power supplies to generate, and it would have to be dissipated every bit that is sent.
More clarification would also be cool!
ekr
Some additional downsides to a bus network:
If there is a break in the main cable, the entire network can go down, with limited to no remaining connectivity.
If something does break down, it can be difficult to identify the cause of the problem.
msebek
@ekr Interesting point. You mention "main cable breaks", which makes sense in the context of distributed systems, where nodes are capable of going on and offline.
Is there any sort of fault tolerance built in to processors? I imagine the chances of a core or the bus being damaged is much lower than the chances of the janitor unplugging the servers while cleaning, etc. I sort of doubt it, since fault tolerance has to be built into the protocol, so would be an overhead for every functioning chip.
Then again, with so many cores on a GPU, the chances that some of them have manufacturing errors are non-negligible. What happens in these cases?
Why electrical load is high?
@sanchuah if I were to hazard a guess, it might be because driving the voltage up or down on the entire bus requires non-trivial current to do so quickly.
This sort of rushing current is inefficient for power supplies to generate, and it would have to be dissipated every bit that is sent.
More clarification would also be cool!
Some additional downsides to a bus network:
@ekr Interesting point. You mention "main cable breaks", which makes sense in the context of distributed systems, where nodes are capable of going on and offline.
Is there any sort of fault tolerance built in to processors? I imagine the chances of a core or the bus being damaged is much lower than the chances of the janitor unplugging the servers while cleaning, etc. I sort of doubt it, since fault tolerance has to be built into the protocol, so would be an overhead for every functioning chip.
Then again, with so many cores on a GPU, the chances that some of them have manufacturing errors are non-negligible. What happens in these cases?