학술논문
In-datacenter performance analysis of a tensor processing unit
Document Type
Conference
Author
Jouppi, Norman P.; Young, Cliff; Patil, Nishant; Patterson, David; Agrawal, Gaurav; Bajwa, Raminder; Bates, Sarah; Bhatia, Suresh; Boden, Nan; Borchers, Al; Boyle, Rick; Cantin, Pierre-luc; Chao, Clifford; Clark, Chris; Coriell, Jeremy; Daley, Mike; Dau, Matt; Dean, Jeffrey; Gelb, Ben; Ghaemmaghami, Tara Vazir; Gottipati, Rajendra; Gulland, William; Hagmann, Robert; Ho, C. Richard; Hogberg, Doug; Hu, John; Hundt, Robert; Hurt, Dan; Ibarz, Julian; Jaffey, Aaron; Jaworski, Alek; Kaplan, Alexander; Khaitan, Harshit; Killebrew, Daniel; Koch, Andy; Kumar, Naveen; Lacy, Steve; Laudon, James; Law, James; Le, Diemthu; Leary, Chris; Liu, Zhuyuan; Lucke, Kyle; Lundin, Alan; MacKean, Gordon; Maggiore, Adriana; Mahony, Maire; Miller, Kieran; Nagarajan, Rahul; Narayanaswami, Ravi; Ni, Ray; Nix, Kathy; Norrie, Thomas; Omernick, Mark; Penukonda, Narayana; Phelps, Andy; Ross, Jonathan; Ross, Matt; Salek, Amir; Samadiani, Emad; Severn, Chris; Sizikov, Gregory; Snelham, Matthew; Souter, Jed; Steinberg, Dan; Swing, Andy; Tan, Mercedes; Thorson, Gregory; Tian, Bo; Toma, Horia; Tuttle, Erick; Vasudevan, Vijay; Walter, Richard; Wang, Walter; Wilcox, Eric; Yoon, Doe Hyun
Source
2017 ACM/IEEE 44th Annual International Symposium on Computer Architecture (ISCA) Computer Architecture (ISCA), 2017 ACM/IEEE 44th Annual International Symposium on. :1-12 Jun, 2017
Subject
Language
Abstract
Many architects believe that major improvements in cost-energy-performance must now come from domain-specific hardware. This paper evaluates a custom ASIC-called a Tensor Processing Unit (TPU)-deployed in datacenters since 2015 that accelerates the inference phase of neural networks (NN). The heart of the TPU is a 65,536 8-bit MAC matrix multiply unit that offers a peak throughput of 92 TeraOps/second (TOPS) and a large (28 MiB) software-managed on-chip memory. The TPU's deterministic execution model is a better match to the 99th-percentile response-time requirement of our NN applications than are the time-varying optimizations of CPUs and GPUs that help average throughput more than guaranteed latency. The lack of such features helps explain why, despite having myriad MACs and a big memory, the TPU is relatively small and low power. We compare the TPU to a server-class Intel Haswell CPU and an Nvidia K80 GPU, which are contemporaries deployed in the same datacenters. Our workload, written in the high-level TensorFlow framework, uses production NN applications (MLPs, CNNs, and LSTMs) that represent 95% of our datacenters' NN inference demand. Despite low utilization for some applications, the TPU is on average about 15X–30X faster than its contemporary GPU or CPU, with TOPS/Watt about 30X–80X higher. Moreover, using the GPU's GDDR5 memory in the TPU would triple achieved TOPS and raise TOPS/Watt to nearly 70X the GPU and 200X the CPU.