In-Datacenter Performance Analysis of a Tensor Processing Unit
a systolic matrix unit, and the roofline that explains it
Read it
- 13 sections
- 0 figures
- 0 equations
- 70 references
- cited by 0 here
- difficulty 2 of 5
- the publisher's page
Contents
- 1Introduction to Neural Networks
- 2TPU Origin, Architecture, and Implementation
- 3CPU, GPU, and TPU Platforms
- 4Performance: Rooflines, Response-Time, and Throughput
- 5Cost-Performance, TCO, and Performance/Watt
- 6Energy Proportionality
- 7Evaluation of Alternative TPU Designs
- 8Discussion
- 9Related Work
- 10Conclusion
- Acknowledgments