Imagine if folks used custom chips instead of standard gaming chips for deep learning. ~100X boost possible...https://twitter.com/mappingbabel/status/676853665500495872 …
@adapteva how did you get to this figure? deep learning applications generally seem to be bound by global memory (off-chip SDRAM) bandwidth.
-
-
-
@nachiketkapre@adapteva Then you'd have to use eg. SGD instead of batch methods like BFGS. Afaik SGD usually converges much much slower.. - Show replies
New conversation -
-
-
@oe1cxw General statement on asics vs gpus. Plenty of room at bottom. For starters...ieee floats and graphics ram huge energy sinks. -
@adapteva half precision ieee floats (aka FP16, supported since CUDA 7.5) are a pretty good match for machine learning applications.
End of conversation
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.