Dave Patterson Evaluation of the Tensor Processing Unit

[@SergiuM]

Microsoft Catapult's FPGA solution was latency optimized for batch size of 1, i.e. it delivered the fastest inference and full utilization without the need to group input samples in a batch. The TPU is also written in Verilog or some RTL, and the FPGA is configured once and used as a soft ASIC, with firmware and SW stack on top.

[@klam77]

SODA hall baby!

[@LydellAaron]

35:11 computer architecture history is important.

[@lawrencechen2004]

46:08 a new DSA, Yeah!

[@foot-book6023]

Is it possible to share the slides from the above presentation?

[@suhasrpai]

drive.google.com/file/d/0B5vMMuEXssYnR1oxcXVaaTFTQWM/view

[@foot-book6023]

thanks man. Got it

[@karthik3685]

45:30 -> wafer scale integration didn't work out. :) Cerbras 2yrs later: hold my wafer :)

[@douggale5962]

Amdahl's law is irritating to hear in talks like this, because it makes a silly assumption that the whole chip is running one program. As if there is "a sequential part" in a sea separate of background crap running simultaneously.

[@pcuimac]

Sounds a bit like comparing apples and oranges. Why would you compare GPUs CPUs and TPUs? They are designed to do completely different things!? Does he want to sell these new chips to big companies? Science and monetary interests don't fit very well together. Good science produces economic growth, but scientists who want to be sconomically successful are often exagerating. We european students oftne found american profs insufferable, because they made every small step into a BIG story about personal success. It's nice to see that he killed the creed that Moores law is still valid. Economists always come up with these "laws" so they don't have to talk about what's next. As long as the dollar is roling they are happy.

[@Piraco]

"Economists always come up with these "laws" so they don't have to talk about what's next" lol

[@137akash]

He is comparing apples to apples in terms of comparing performance and power of CPU/GPU/TPU with respect to their consumption while computing Neural Networks. In High Throughput computing required for Neural Networks TPU works better than GPU/CPU. This guy is from academics and he doesn't have to worry about company(Google) economics or marketing!!!

[@FredIsMyName22]

what a stupid comment lol.