3 ms·
It's something lots of research groups experiment with, but apparently it doesn't pay off all that often. Working with large memory datasets is something most
by detaro 6y ago
It's something lots of research groups experiment with, but apparently it doesn't pay off all that often.
Working with large memory datasets is something most accelerators aren't good at, because for most their memory access speed outside a limited local set of data is limited, so they can't easily outcompete traditional optimizations/scaling. It can make sense for some cases, e.g. there are some few GPU-accelerated databases, and you can spend lots of money on fancy FPGA accelerators, but the competition is always "why not just use a bigger/more traditional servers instead of exotic special things".
- matesz 6y agoThanks for your answer - in a nutshell big companies do consider it, but decide to wait for chip manufacturers to update their hardware and do the upgrade. How fast is this feedback loop matters here.