Many architects believe that major improvements in cost-energyperformance 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 responsetime 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.

The performance tradeoff between hardware complexity and clock speed is studied. First, a generic superscalar pipeline is defined. Then the specific areas of register renaming, instruction window wakeup and selection logic, and operand bypassing are analyzed. Each is modeled and Spice simulated for feature sizes of 0:8m, 0:35m, and 0:18m. Performance results and trends are expressed in terms of issue width and window size. Our analysis indicates that window wakeup and selection logic as well as operand bypass logic are likely to be the most critical in the future.A microarchitecture that simplifies wakeup and selection logic is proposed and discussed. This implementation puts chains of dependent instructions into queues, and issues instructions from multiple queues in parallel. Simulation shows little slowdown as compared with a completely flexible issue window when performance is measured in clock cycles. Furthermore, because only instructions at queue heads need to be awakened and selected, issue logic is simplified and the clock cycle is faster -consequently overall performance is improved. By grouping dependent instructions together, the proposed microarchitecture will help minimize performance degradation due to slow bypasses in future wide-issue machines.

This paper describes an analytical model for the access and cycle times of on-chip directmapped and set-associative caches. The inputs to the model are the cache size, block size, and associativity, a s w ell as array organization and process parameters. The model gives estimates that are within 6 of Hspice results for the circuits we h a ve c hosen. This model extends previous models and xes many of their major shortcomings. New features include models for the tag array, comparator, and multiplexor drivers, non-step stage input slopes, rectangular stacking of memory subarrays, a transistor-level decoder model, column-multiplexed bitlines controlled by an additional array organizational parameter, load-dependent size transistors for wordline drivers, and output of cycle times as well as access times. Software implementing the model is available via ftp.

Abstract

Many architects believe that major improvements in cost-energyperformance 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 responsetime 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.