In-datacenter performance analysis of a tensor processing unit
By Norman P. Jouppi, Cliff Young, Nishant Patil, David A. Patterson, Gaurav Agrawal, R. Bajwa, S. Bates, S. Bhatia, Nan Boden, N. Boden and 70 more. 2017
5,323 citations recorded by the source corpus.
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.
Subfields: Computer Networks and Communications, Electrical and Electronic Engineering, Hardware and Architecture, ML Systems & Hardware
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Computer Networks and Communications
- Medhi Bennis
- Ion Stoica
- Merouane Debbah
- Walid Saad
- Jihong Park
- Carole-Jean Wu
- Onur Mutlu
- Pieter Abbeel
In Electrical and Electronic Engineering
- Shimeng Yu
- Daniele Ielmini
- Medhi Bennis
- Song Han
- Merouane Debbah
- H. Philip Wong
- Walid Saad
- Carole-Jean Wu
In Hardware and Architecture
In ML Systems & Hardware
- Carole-Jean Wu
- Shimeng Yu
- Young Geun Kim
- Medhi Bennis
- Ion Stoica
- Quoc V. Le
- Tien-Ju Yang
- Vivienne Sze
Citations in the corpus
- Training Compute-Optimal Large Language Models (2022, neutral): “All models in this analysis have been trained on TPUv3/TPUv4 (Jouppi et al., 2017) with JAX (Bradbury et al., 2018) and Haiku (Hennigan et al., 2020).”
- Massively Multilingual Neural Machine Translation in the Wild: Findings and Challenges (2019, methodology): “baselines we use a batch size of 1M tokens perbatch, by using large scale data parallelism over 16 TPUv3 chips (Jouppi et al., 2017).”
- Memory-Efficient Adaptive Optimization for Large-Scale Learning (2019, methodology): “In all experiments, we used the Cloud TPU-v2 device [14] where each core has 8GiB of memory.”
- Measuring the Effects of Data Parallelism on Neural Network Training (2018, methodology): “lism in the synchronous training setting. Hardware for training neural networks is trending towards ever-increasing capacity for data parallelism. Specialized systems using GPUs or custom ASICs (e.g. Jouppi et al., 2017) combined with high-performance interconnect technology are unlocking unpreceden”
- GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding (2020, background): “More recently, the industry also invested heavily in building more dedicated hardware systems chasing for more cost-effective neural network hardware (Jouppi et al., 2017).”
- Memory-Efficient Pipeline-Parallel DNN Training (2020, background): “For example, a deeper configuration has additional memory capacity, allowing for a larger maximum per-GPU microbatch size; this can increase the arithmetic intensity (number of floating point operations performed per memory load) of kernels (Jouppi et al., 2017), and consequently throughput.”
- Don't Decay the Learning Rate, Increase the Batch Size (2017, methodology): “UTES To confirm that increasing the batch size during training can reduce model training times, we replicated the set-up described by Goyal et al. (2017) on a half TPU pod, comprising 256 tensorcores (Jouppi et al., 2017). Using tensorFlow, we first train ResNet-50 for 90 epochs to 76.1% validation se”
- FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (2022, background): “[46] Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, et al.”
- Sustainable AI: Environmental Implications, Challenges and Opportunities (2021, background): “Existing practices such as the move to domain-specific architectures at cloud scale [68], [69], [70] reduce AI computing’s footprint by consolidating computing resources at scale and by operating the shared infrastructures more environmentally-friendly with carbon free energy6.”
- Big Transfer (BiT): General Visual Representation Learning (2019, methodology): “ochs with an initial learning rate of 0.03, with 5000 linear warmup steps and decay the learning rate after 10, 23, 30 and 37 epochs. We use a global batch size of 4096 and train on a Cloud TPUv3-512 [22], resulting in 8 images per chip. For optimization with large batch sizes we employ recipes from”
- Perceiver IO: A General Architecture for Structured Inputs & Outputs (2021, methodology): “We use a larger batch size of 8192 and train on 256 TPUs.”
- VideoGPT: Video Generation using VQ-VAE and Transformers (2021, neutral): “…learning architectures (He et al., 2015; van den Oord et al., 2016b;c; Vaswani et al., 2017; Zhang et al., 2019; Menick & Kalch-brenner, 2018) as well as the availability of compute resources (Jouppi et al., 2017; Amodei & Hernandez, 2018) that are more powerful and plentiful than a few years ago.”
- Revisiting Self-Supervised Visual Representation Learning (2019, methodology): “We thank Sylvain Gelly for many fruitful discussions and Marvin Ritter for helping us to run experiments using TPUs.”
- Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm (2017, methodology): “Training proceeded for 700,000 steps (mini-batches of size 4,096) starting from randomly initialised parameters, using 5,000 first-generation TPUs (16) to generate self-play games and 64 second-generation TPUs to train the neural networks; training lasted for approximately 9 hours in chess, 12 hours”