Search

Quantization has become a mainstream compression technique for reducing model size, computational requirements, and energy consumption for modern deep neural networks (DNNs). With improved numerical support in recent hardware, including multiple variants of integer and floating point, mixed-precision quantization has become necessary to achieve high-quality results with low model cost. Prior mixed-precision methods have performed either a post-training quantization search, which compromises on accuracy, or a differentiable quantization search, which leads to high memory usage from branching. Therefore, we propose the first one-shot mixed-precision quantization search that eliminates the need for retraining in both integer and low-precision floating point models. We evaluate our search (FLIQS) on multiple convolutional and vision transformer networks to discover Pareto-optimal models. Our approach improves upon uniform precision, manual mixed-precision, and recent integer quantization search methods. With integer models, we increase the accuracy of ResNet-18 on ImageNet by 1.31% and ResNet-50 by 0.90% with equivalent model cost over previous methods. Additionally, for the first time, we explore a novel mixed-precision floating-point search and improve MobileNetV2 by up to 0.98% compared to prior state-of-the-art FP8 models. Finally, we extend FLIQS to simultaneously search a joint quantization and neural architecture space and improve the ImageNet accuracy by 2.69% with similar model cost on a MobileNetV2 search space., Comment: Accepted to AutoML 2024

We expect that many-core microprocessors will push performance per chip from the 10 gigaflop to the 10 teraflop range in the coming decade. To support this increased performance, memory and inter-core bandwidths will also have to scale by orders of magnitude. Pin limitations, the energy cost of electrical signaling, and the non-scalability of chip-length global wires are significant bandwidth impediments. Recent developments in silicon nanophotonic technology have the potential to meet these off- and on- stack bandwidth requirements at acceptable power levels. Corona is a 3D many-core architecture that uses nanophotonic communication for both inter-core communication and off-stack communication to memory or I/O devices. Its peak floating-point performance is 10 teraflops. Dense wavelength division multiplexed optically connected memory modules provide 10 terabyte per second memory bandwidth. A photonic crossbar fully interconnects its 256 low-power multithreaded cores at 20 terabyte per second bandwidth. We have simulated a 1024 thread Corona system running synthetic benchmarks and scaled versions of the SPLASH-2 benchmark suite. We believe that in comparison with an electrically-connected many-core alternative that uses the same on-stack interconnect power, Corona can provide 2 to 6 times more performance on many memory-intensive workloads, while simultaneously reducing power., Comment: This edition is recompiled from proceedings of ISCA-35 (the 35th International Symposium on Computer Architecture, June 21 - 25, 2008, Beijing, China) and has minor formatting differences. 13 pages; 11 figures

The 2008 Corona effort was inspired by a pressing need for more of everything, as demanded by the salient problems of the day. Dennard scaling was no longer in effect. A lot of computer architecture research was in the doldrums. Papers often showed incremental subsystem performance improvements, but at incommensurate cost and complexity. The many-core era was moving rapidly, and the approach with many simpler cores was at odds with the better and more complex subsystem publications of the day. Core counts were doubling every 18 months, while per-pin bandwidth was expected to double, at best, over the next decade. Memory bandwidth and capacity had to increase to keep pace with ever more powerful multi-core processors. With increasing core counts per die, inter-core communication bandwidth and latency became more important. At the same time, the area and power of electrical networks-on-chip were increasingly problematic: To be reliably received, any signal that traverses a wire spanning a full reticle-sized die would need significant equalization, re-timing, and multiple clock cycles. This additional time, area, and power was the crux of the concern, and things looked to get worse in the future. Silicon nanophotonics was of particular interest and seemed to be improving rapidly. This led us to consider taking advantage of 3D packaging, where one die in the 3D stack would be a photonic network layer. Our focus was on a system that could be built about a decade out. Thus, we tried to predict how the technologies and the system performance requirements would converge in about 2018. Corona was the result this exercise; now, 15 years later, it's interesting to look back at the effort., Comment: 2 pages. Proceedings of ISCA-50: 50 years of the International Symposia on Computer Architecture (selected papers) June 17-21 Orlando, Florida

In response to innovations in machine learning (ML) models, production workloads changed radically and rapidly. TPU v4 is the fifth Google domain specific architecture (DSA) and its third supercomputer for such ML models. Optical circuit switches (OCSes) dynamically reconfigure its interconnect topology to improve scale, availability, utilization, modularity, deployment, security, power, and performance; users can pick a twisted 3D torus topology if desired. Much cheaper, lower power, and faster than Infiniband, OCSes and underlying optical components are <5% of system cost and <3% of system power. Each TPU v4 includes SparseCores, dataflow processors that accelerate models that rely on embeddings by 5x-7x yet use only 5% of die area and power. Deployed since 2020, TPU v4 outperforms TPU v3 by 2.1x and improves performance/Watt by 2.7x. The TPU v4 supercomputer is 4x larger at 4096 chips and thus ~10x faster overall, which along with OCS flexibility helps large language models. For similar sized systems, it is ~4.3x-4.5x faster than the Graphcore IPU Bow and is 1.2x-1.7x faster and uses 1.3x-1.9x less power than the Nvidia A100. TPU v4s inside the energy-optimized warehouse scale computers of Google Cloud use ~3x less energy and produce ~20x less CO2e than contemporary DSAs in a typical on-premise data center., Comment: 15 pages; 16 figures; to be published at ISCA 2023 (the International Symposium on Computer Architecture)

Neural Architecture Search (NAS), together with model scaling, has shown remarkable progress in designing high accuracy and fast convolutional architecture families. However, as neither NAS nor model scaling considers sufficient hardware architecture details, they do not take full advantage of the emerging datacenter (DC) accelerators. In this paper, we search for fast and accurate CNN model families for efficient inference on DC accelerators. We first analyze DC accelerators and find that existing CNNs suffer from insufficient operational intensity, parallelism, and execution efficiency. These insights let us create a DC-accelerator-optimized search space, with space-to-depth, space-to-batch, hybrid fused convolution structures with vanilla and depthwise convolutions, and block-wise activation functions. On top of our DC accelerator optimized neural architecture search space, we further propose a latency-aware compound scaling (LACS), the first multi-objective compound scaling method optimizing both accuracy and latency. Our LACS discovers that network depth should grow much faster than image size and network width, which is quite different from previous compound scaling results. With the new search space and LACS, our search and scaling on datacenter accelerators results in a new model series named EfficientNet-X. EfficientNet-X is up to more than 2X faster than EfficientNet (a model series with state-of-the-art trade-off on FLOPs and accuracy) on TPUv3 and GPUv100, with comparable accuracy. EfficientNet-X is also up to 7X faster than recent RegNet and ResNeSt on TPUv3 and GPUv100.

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 (caches, out-of-order execution, multithreading, multiprocessing, prefetching, ...) 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., Comment: 17 pages, 11 figures, 8 tables. To appear at the 44th International Symposium on Computer Architecture (ISCA), Toronto, Canada, June 24-28, 2017