Post a Comment
The article is short in information, but I would assume those laser interconnections would be used for internal RAM(registers, caches, whatever) as well. Even if the input/output speed between this machine and external conventional computers is relatively slow, you would only need to load the data once, because this massively parallelized machine isn't intended for serial calculations anyways, I guess.
You're right. Parallel calculations don't work that way. They tend to pass large lumps of data around in one go to perform things like matrix calculations and reductions on the data set. PCIe and a decent interconnect can do a pretty good job of shoving that data out of the local host memory and getting it to another node pretty quickly. Decent switches and things like multi-casting can help considerably.
Generally though the trick isn't finding enough bandwidth, it's reducing latency.








