So why *aren't* there small bits of compute baked into RAM yet? Referring to highly local operations with fixed dimensionality, or reducing queries (return or update memory region where following constraint is true). We have TB+ EC2 nodes, after all. Why move the data?
The thing is DRAM doesn't spend a huge fraction of the time refreshing itself. DRAM is pipelined and you can stream data in/out of RAM pretty much as fast as the physical technology allows.
-
-
Exactly! The streaming of data in and out _is the bottleneck_. So add compute when you are already processing the data anyway.
-
I'm not convinced you could put enough useful compute in RAM chips to be useful. Sure, streaming workloads would benefit, but those often involve big fat SIMD instructions. And then you have to deal with cache coherency.
- Show replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.