Multi-pod systolic arrays are emerging as the architecture of choice in DNN inference accelerators. Despite their potential, designing multi-pod systolic arrays to maximize effective throughput/Watt—i.e., throughput/Watt adjusted when accounting for array utilization—poses a unique set of challenges. In this work, we study three key pillars in multi-pod systolic array designs, namely array granularity, interconnect, and tiling. We identify optimal array granularity across workloads and show that state-of-the-art commercial accelerators use suboptimal array sizes for single-tenancy workloads. We, then evaluate the bandwidth/latency trade-offs in interconnects and show that Butterfly networks offer a scalable topology for accelerators with a large number of pods. Finally, we introduce a novel data tiling scheme with custom partition size to maximize utilization in optimally sized pods. We propose Scale-out Systolic Arrays , a multi-pod inference accelerator for both single- and multi-tenancy based on these three pillars. We show that SOSA exhibits scaling of up to 600 TeraOps/s in effective throughput for state-of-the-art DNN inference workloads, and outperforms state-of-the-art multi-pod accelerators by a factor of 1.5 ×. 1

DNN inference accelerators executing online services exhibit low average loads because of service demand variability, leading to poor resource utilization. Unfortunately, reclaiming idle inference cycles is difficult as other workloads can not execute on a custom accelerator. With recent proposals for the use of fixed-point arithmetic in training, there are opportunities for training services to piggyback on inference accelerators. We make the observation that a key challenge in doing so is maintaining service-level latency constraints for inference. We show that relaxing latency constraints in an inference accelerator with ALU arrays that are batching-optimized achieves near-optimal throughput for a given area and power envelope while maintaining inference services’ tail latency goals.

Stencil codes (a.k.a. nearest-neighbor computations) are widely used in image processing, machine learning, and scientific applications. Stencil codes incur nearest-neighbor data exchange because the value of each point in the structured grid is calculated as a function of its value and the values of a subset of its nearest-neighbor points. When running on Graphics Processing Unit (GPUs), stencil codes exhibit a high degree of data sharing between nearest-neighbor threads. Sharing is typically implemented through shared memories, shuffle instructions, and on-chip caches and often incurs performance overheads due to the redundancy in memory accesses. In this article, we propose Neighbor Data (NeDa), a direct nearest-neighbor data sharing mechanism that uses two registers embedded in each streaming processor (SP) that can be accessed by nearest-neighbor SP cores. The registers are compiler-allocated and serve as a data exchange mechanism to eliminate nearest-neighbor shared accesses. NeDa is embedded carefully with local wires between SP cores so as to minimize the impact on density. We place and route NeDa in an open-source GPU and show a small area overhead of 1.3%. The cycle-accurate simulation indicates an average performance improvement of 21.8% and power reduction of up to 18.3% for stencil codes in General-Purpose Graphics Processing Unit (GPGPU) standard benchmark suites. We show that NeDa’s performance is within 13.2% of an ideal GPU with no overhead for nearest-neighbor data exchange.

Modern online services are shifting away from monolithic applications to loosely-coupled microservices because of their improved scalability, reliability, programmability and development velocity. Microservices communicating over the datacenter network require data transformation (DT) to convert messages back and forth between their internal formats. This work identifies DT as a bottleneck due to reductions in latency of the surrounding system components, namely application runtimes, protocol stacks, and network hardware. We therefore propose Optimus Prime (OP), a programmable DT accelerator that uses a novel abstraction, an in-memory schema, to represent DT operations. The schema is compatible with today's DT frameworks and enables any compliant accelerator to perform the transformations comprising a request in parallel. Our evaluation shows that OP's DT throughput matches the line rate of today's NICs and has ~60x higher throughput compared to software, at a tiny fraction of the CPU's silicon area and power. We also evaluate a set of microservices running on Thrift, and show up to 30% reduction in service latency.

Graphics Processing Units (GPUs) employ large register files to accommodate all active threads and accelerate context switching. Unfortunately, register files are a scalability bottleneck for future GPUs due to long access latency, high power consumption, and large silicon area provisioning. Prior work proposes hierarchical register file to reduce the register file power consumption by caching registers in a smaller register file cache. Unfortunately, this approach does not improve register access latency due to the low hit rate in the register file cache. In this article, we propose the Latency-Tolerant Register File (LTRF) architecture to achieve low latency in a two-level hierarchical structure while keeping power consumption low. We observe that compile-time interval analysis enables us to divide GPU program execution into intervals with an accurate estimate of a warp’s aggregate register working-set within each interval. The key idea of LTRF is to prefetch the estimated register working-set from the main register file to the register file cache under software control, at the beginning of each interval, and overlap the prefetch latency with the execution of other warps. We observe that register bank conflicts while prefetching the registers could greatly reduce the effectiveness of LTRF. Therefore, we devise a compile-time register renumbering technique to reduce the likelihood of register bank conflicts. Our experimental results show that LTRF enables high-capacity yet long-latency main GPU register files, paving the way for various optimizations. As an example optimization, we implement the main register file with emerging high-density high-latency memory technologies, enabling 8× larger capacity and improving overall GPU performance by 34%.

Graphics Processing Units (GPUs) employ large register files to accommodate all active threads and accelerate context switching. Unfortunately, register files are a scalability bottleneck for future GPUs due to long access latency, high power consumption, and large silicon area provisioning. Prior work proposes hierarchical register file, to reduce the register file power consumption by caching registers in a smaller register file cache. Unfortunately, this approach does not improve register access latency due to the low hit rate in the register file cache. In this paper, we propose the Latency-Tolerant Register File (LTRF) architecture to achieve low latency in a two-level hierarchical structure while keeping power consumption low. We observe that compile-time interval analysis enables us to divide GPU program execution into intervals with an accurate estimate of a warp's aggregate register working-set within each interval. The key idea of LTRF is to prefetch the estimated register working-set from the main register file to the register file cache under software control, at the beginning of each interval, and overlap the prefetch latency with the execution of other warps. Our experimental results show that LTRF enables high-capacity yet long-latency main GPU register files, paving the way for various optimizations. As an example optimization, we implement the main register file with emerging high-density high-latency memory technologies, enabling 8X larger capacity and improving overall GPU performance by 31% while reducing register file power consumption by 46%.

The wide adoption of DNNs has given birth to unrelenting computing requirements, forcing datacenter operators to adopt domain-specific accelerators to train them. These accelerators typically employ densely packed full-precision floating-point arithmetic to maximize performance per area. Ongoing research efforts seek to further increase that performance density by replacing floating-point with fixed-point arithmetic. However, a significant roadblock for these attempts has been fixed point's narrow dynamic range, which is insufficient for DNN training convergence. We identify block floating point (BFP) as a promising alternative representation since it exhibits wide dynamic range and enables the majority of DNN operations to be performed with fixed-point logic. Unfortunately, BFP alone introduces several limitations that preclude its direct applicability. In this work, we introduce HBFP, a hybrid BFP-FP approach, which performs all dot products in BFP and other operations in floating point. HBFP delivers the best of both worlds: the high accuracy of floating point at the superior hardware density of fixed point. For a wide variety of models, we show that HBFP matches floating point's accuracy while enabling hardware implementations that deliver up to 8.5x higher throughput.

Graphic processing unit (GPU)-based general-purpose computing is developing as a viable alternative to CPU-based computing in many domains. Today’s tools for GPU analysis include simulators like GPGPU-Sim, Multi2Sim, and Barra. While useful for modeling first-order effects, these tools do not provide a detailed view of GPU microarchitecture and physical design. Further, as GPGPU research evolves, design ideas and modifications demand detailed estimates of impact on overall area and power. Fueled by this need, we introduce MIAOW (Many-core Integrated Accelerator Of Wisconsin), an open-source RTL implementation of the AMD Southern Islands GPGPU ISA, capable of running unmodified OpenCL-based applications. We present our design motivated by our goals to create a realistic, flexible, OpenCL-compatible GPGPU, capable of emulating a full system. We first explore if MIAOW is realistic and then use four case studies to show that MIAOW enables the following: physical design perspective to “traditional” microarchitecture, new types of research exploration, and validation/calibration of simulator-based characterization of hardware. The findings and ideas are contributions in their own right, in addition to MIAOW’s utility as a tool for others’ research.