3 ms·
It will be interesting to see how this compares to Google's custom TPU: https://cloudplatform.googleblog.com/2016/05/Google-supercharges-machine-learning-tasks-
by ucaetano 10y ago
It will be interesting to see how this compares to Google's custom TPU:
https://cloudplatform.googleblog.com/2016/05/Google-supercharges-machine-learning-tasks-with-custom-chip.html https://cloudplatform.googleblog.com/2016/05/Google-supercha...
NVidia is still taking the one-size-fits-all approach to AI and graphics, maybe it is time to develop AI-specific hardware.
- melling 10y agoIsn't this simple economics? The R&D cost is amortized over a larger market. You can always try to build specialized chips but the market might not be nearly as big so there's a lot less money. The general purpose market still moves faster.
- ucaetano 10y agoGiven that the market is big enough for a single player to develop a chip for internal usage, minimum efficient scale is probably not an issue. I'd guess NVidia sells far more cards than Google will produce TPUs.