A whole butt ton of GPU-style FMA and low precision float multiply ALUs would be... | Hacker News

Hacker Newsnew | past | comments | ask | show | jobs | submit

		hatsunearu on July 24, 2017 \| parent \| context \| favorite \| on: Second version of HoloLens HPU will incorporate AI... A whole butt ton of GPU-style FMA and low precision float multiply ALUs would be my guess

protomyth on July 24, 2017 [–]

How low precision can you get and still have it be useful?

fulafel on July 24, 2017 | | [–]

Google's TPU proves that 8 bits is still good.

sbierwagen on July 24, 2017 | | | [–]

Note that the TPU used 8 bit integer math, not even floating point.

Govindae on July 24, 2017 | | [–]

1-bit, binary neural networks work.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact