Hacker News
new
|
past
|
comments
|
ask
|
show
|
jobs
|
submit
login
astrodust
on July 22, 2016
|
parent
|
context
|
favorite
| on:
Nvidia CEO Reveals New TITAN X at Stanford Deep L...
You can get away with 8-bit floats on many neural network type applications so that might be the idea here.
bitfiddler0
on July 22, 2016
[–]
It's an 8-bit int type though. Perhaps even that's sufficient for inference?
astrodust
on July 23, 2016
|
parent
[–]
For a lot of network types that will do the job. 2x the performance is often better than 2x the fidelity. You can get higher accuracy with more nodes.
Consider applying for YC's Winter 2027 batch!
Applications
are open till November 2.
Guidelines
|
FAQ
|
Lists
|
API
|
Security
|
Legal
|
Apply to YC
|
Contact
Search:
