Conversation
Notices
-
Embed this notice
Account: Computers (pro@mu.zaitcev.nu)'s status on Thursday, 30-Nov-2023 01:42:37 JST Account: Computers
https://www.beagleboard.org/boards/beaglev-ahead
This has something called "Neural Processing Unit". What does that do? What is it for?-
Embed this notice
Sexy Moon (moon@shitposter.club)'s status on Thursday, 30-Nov-2023 01:42:36 JST Sexy Moon
@pro > NPUs are capable of handling machine learning workloads as much as 10,000 times faster than a standard GPU, and they’re also more power-efficient[.] In conversation permalink -
Embed this notice
Sexy Moon (moon@shitposter.club)'s status on Thursday, 30-Nov-2023 01:47:20 JST Sexy Moon
@pro I had to learn that int8 is the quantization model data type and 4TOPS is a cross-platform metric. I still don't get it though. In conversation permalink -
Embed this notice
Sexy Moon (moon@shitposter.club)'s status on Thursday, 30-Nov-2023 01:47:46 JST Sexy Moon
@pro Google Coral TPU is apparently also 4TOPS INT8. In conversation permalink -
Embed this notice
Account: Computers (pro@mu.zaitcev.nu)'s status on Thursday, 30-Nov-2023 02:07:19 JST Account: Computers
@Moon I see, thanks. So basically it is a hardware for inference. In conversation permalink Sexy Moon likes this. -
Embed this notice
Sexy Moon (moon@shitposter.club)'s status on Thursday, 30-Nov-2023 02:21:06 JST Sexy Moon
@pro training is making the model. inference is applying the model to generate output. in terms of processing I don't know how they differ either. I think these are all inference-only. In conversation permalink -
Embed this notice
Account: Computers (pro@mu.zaitcev.nu)'s status on Thursday, 30-Nov-2023 02:21:09 JST Account: Computers
@Moon I forgot to add the thank you meme. In reality, I have no idea how inference differs from training, but whatever. I was just curious. In conversation permalink Attachments
Sexy Moon likes this.
-
Embed this notice