Conversation
|
things im a little iffy on
|
|
I trolled and I thought the bad cores were the high index ones... quantized the embedding model to fp16, tinygrad LLVM has worse performance than fp32 on the bad core, so switched backend to CPU=1 and gets 42% CPU usage on bad cores (however on the good core, LLVM seems to do better than CPU with the quantized model???) there also seems to be possibly be some performance gap depending on if the net was compiled on the same type of core it is run on |
|
17-18% little core CPU usage |
|
IMAGE=2 compiles the model to be ~ same time as BEAM=1 so this the previous cpu/gpu usage is the same, and compile times are normal
|
adeebshihadeh
left a comment
There was a problem hiding this comment.
Did a quick pass, looking into openWakeWord now
|
< 8% little core CPU usage... unsure how usable this is model doesn't seem to get much better by scaling model size... perhaps real user data will fix it? |
|
This PR has had no activity for 9 days. It will be automatically closed in 2 days if there is no activity. |
Saying "Hey comma" will let you bookmark a segment, in the same way that clicking the bookmark button in the sidebar currently does.