Computer Science

Recent advancements in deep learning have led to the widespread adoption of artificial intelligence (AI) in applications such as computer vision and natural language processing. As neural networks become deeper and larger, AI modeling demands outstrip the capabilities of conventional chip architectures. Memory bandwidth falls behind processing power. Energy consumption comes to dominate the total cost of ownership. Currently, memory capacity is insufficient to support the most advanced NLP models. In this work, we present a 3D AI chip, called Sunrise, with near-memory computing architecture to address these three challenges. This distributed, near-memory computing architecture allows us to tear down the performance-limiting memory wall with an abundance of data bandwidth. We achieve the same level of energy efficiency on 40nm technology as competing chips on 7nm technology. By moving to similar technologies as other AI chips, we project to achieve more than ten times the energy efficiency, seven times the performance of the current state-of-the-art chips, and twenty times of memory capacity as compared with the best chip in each benchmark.

The emergence of P4, a domain specific language, coupled to PISA, a domain specific architecture, is revolutionizing the networking field. P4 allows to describe how packets are processed by a programmable data plane, spanning ASICs and CPUs, implementing PISA. Because the processing flexibility can be limited on ASICs, while the CPUs performance for networking tasks lag behind, recent works have proposed to implement PISA on FPGAs. However, little effort has been dedicated to analyze whether FPGAs are good candidates to implement PISA. In this work, we take a step back and evaluate the micro-architecture efficiency of various PISA blocks. We demonstrate, supported by a theoretical and experimental analysis, that the performance of a few PISA blocks is severely limited by the current FPGA architectures. Specifically, we show that match tables and programmable packet schedulers represent the main performance bottlenecks for FPGA-based programmable switches. Thus, we explore two avenues to alleviate these shortcomings. First, we identify network applications well tailored to current FPGAs. Second, to support a wider range of networking applications, we propose modifications to the FPGA architectures which can also be of interest out of the networking field.

As new technologies are invented, their commercial viability needs to be carefully examined along with their technical merits and demerits. The posit data format, proposed as a drop-in replacement for IEEE 754 float format, is one such invention that requires extensive theoretical and experimental study to identify products that can benefit from the advantages of posits for specific market segments. In this paper, we present an extensive empirical study of posit-based arithmetic vis-?-vis IEEE 754 compliant arithmetic for the optical flow estimation method called Lucas-Kanade (LuKa). First, we use SoftPosit and SoftFloat format emulators to perform an empirical error analysis of the LuKa method. Our study shows that the average error in LuKa with SoftPosit is an order of magnitude lower than LuKa with SoftFloat. We then present the integration of the hardware implementation of a posit adder and multiplier in a RISC-V open-source platform. We make several recommendations, along with the analysis of LuKa in the RISC-V context, for future generation platforms incorporating posit arithmetic units.

GPUs offer orders-of-magnitude higher memory bandwidth than traditional CPU-only systems. However, GPU device memory tends to be relatively small and the memory capacity can not be increased by the user. This paper describes Buddy Compression, a scheme to increase both the effective GPU memory capacity and bandwidth while avoiding the downsides of conventional memory-expanding strategies. Buddy Compression compresses GPU memory, splitting each compressed memory entry between high-speed device memory and a slower-but-larger disaggregated memory pool (or system memory). Highly-compressible memory entries can thus be accessed completely from device memory, while incompressible entries source their data using both on and off-device accesses. Increasing the effective GPU memory capacity enables us to run larger-memory-footprint HPC workloads and larger batch-sizes or models for DL workloads than current memory capacities would allow. We show that our solution achieves an average compression ratio of 2.2x on HPC workloads and 1.5x on DL workloads, with a slowdown of just 1~2%.

Memory controller scheduling is crucial in multicore processors, where DRAM bandwidth is shared. Since increased number of requests from multiple cores of processors becomes a source of bottleneck, scheduling the requests efficiently is necessary to utilize all the computing power these processors offer. However, current multicore processors are using traditional memory controllers, which are designed for single-core processors. They are unable to adapt to changing characteristics of memory workloads that run simultaneously on multiple cores. Existing schedulers may disrupt locality and bank parallelism among data requests coming from different cores. Hence, novel memory controllers that consider and adapt to the memory access characteristics, and share memory resources efficiently and fairly are necessary. We introduce Core-Aware Dynamic Scheduler (CADS) for multicore memory controller. CADS uses Reinforcement Learning (RL) to alter its scheduling strategy dynamically at runtime. Our scheduler utilizes locality among data requests from multiple cores and exploits parallelism in accessing multiple banks of DRAM. CADS is also able to share the DRAM while guaranteeing fairness to all cores accessing memory. Using CADS policy, we achieve 20% better cycles per instruction (CPI) in running memory intensive and compute intensive PARSEC parallel benchmarks simultaneously, and 16% better CPI with SPEC 2006 benchmarks.

The emergence of Phase-Change Memory (PCM) provides opportunities for directly connecting persistent memory to main memory bus. While PCM achieves high read throughput and low standby power, the critical concerns are its poor write performance and limited durability, especially when compared to DRAM. A naturally inspired design is the hybrid memory architecture that fuses DRAM and PCM, so as to exploit the positive aspects of both types of memory. Unfortunately, existing solutions are seriously challenged by the limited main memory size, which is the primary bottleneck of in-memory computing. In this paper, we introduce a novel Content Aware hybrid PCM/DRAM main memory system framework - CARAM, which exploits deduplication to improve line sharing with high memory efficiency. CARAM effectively reduces write traffic to hybrid memory by removing unnecessary duplicate line writes. It also substantially extends available free memory space by coalescing redundant lines in hybrid memory, thereby further improving the wear-leveling efficiency of PCM. To obtain high data access performance, we also design a set of acceleration techniques to minimize the overhead caused by extra computation costs. Our experiment results show that CARAM effectively reduces 15%~42% of memory usage and improves I/O bandwidth by 13%~116%, while saving 31%~38% energy consumption, compared to the state-of-the-art of hybrid systems.

Network bound applications, like a database server executing OLTP queries or a caching server storing objects for a dynamic web applications, are essential services that consumers and businesses use daily. These services run on a large datacenters and are required to meet predefined Service Level Objectives (SLO), or latency targets, with high probability. Thus, efficient datacenter applications should optimize their execution in terms of power and performance. However, to support large scale data storage, these workloads make heavy use of pointer connected data structures (e.g., hash table, large fan-out tree, trie) and exhibit poor instruction and memory level parallelism. Our experiments show that due to long memory access latency, these workloads occupy processor resources (e.g., ROB entries, RS buffers, LS queue entries etc.) for a prolonged period of time that delay the processing of subsequent requests. Delayed execution not only increases request processing latency, but also severely effects an application throughput and power-efficiency. To overcome this limitation, we present CARGO, a novel mechanism to overlap queuing latency and request processing by executing select instructions on an application critical path at the network interface card (NIC) while requests wait for processor resources to become available. Our mechanism dynamically identifies the critical instructions and includes the register state needed to compute the long latency memory accesses. This context-augmented critical region is often executed at the NIC well before execution begins at the core, effectively prefetching the data ahead of time. Across a variety of interactive datacenter applications, our proposal improves latency, throughput, and power efficiency by 2.7X, 2.7X, and 1.5X, respectively, while incurring a modest amount storage overhead.

Convolutional Neural Networks (CNNs) have proven to be extremely accurate for image recognition, even outperforming human recognition capability. When deployed on battery-powered mobile devices, efficient computer architectures are required to enable fast and energy-efficient computation of costly convolution operations. Despite recent advances in hardware accelerator design for CNNs, two major problems have not yet been addressed effectively, particularly when the convolution layers have highly diverse structures: (1) minimizing energy-hungry off-chip DRAM data movements; (2) maximizing the utilization factor of processing resources to perform convolutions. This work thus proposes an energy-efficient architecture equipped with several optimized dataflows to support the structural diversity of modern CNNs. The proposed approach is evaluated by implementing convolutional layers of VGGNet-16 and ResNet-50. Results show that the architecture achieves a Processing Element (PE) utilization factor of 98% for the majority of 3x3 and 1x1 convolutional layers, while limiting latency to 396.9 ms and 92.7 ms when performing convolutional layers of VGGNet-16 and ResNet-50, respectively. In addition, the proposed architecture benefits from the structured sparsity in ResNet-50 to reduce the latency to 42.5 ms when half of the channels are pruned.

Reducing the average memory access time is crucial for improving the performance of applications running on multi-core architectures. With workload consolidation this becomes increasingly challenging due to shared resource contention. Techniques for partitioning of shared resources - cache and bandwidth - and prefetching throttling have been proposed to mitigate contention and reduce the average memory access time. However, existing proposals only employ a single or a subset of these techniques and are therefore not able to exploit the full potential of coordinated management of cache, bandwidth and prefetching. Our characterization results show that application performance, in several cases, is sensitive to prefetching, cache and bandwidth allocation. Furthermore, the results show that managing these together provides higher performance potential during workload consolidation as it enables more resource trade-offs. In this paper, we propose CBP a coordination mechanism for dynamically managing prefetching throttling, cache and bandwidth partitioning, in order to reduce average memory access time and improve performance. CBP works by employing individual resource managers to determine the appropriate setting for each resource and a coordinating mechanism in order to enable inter-resource trade-offs. Our evaluation on a 16-core CMP shows that CBP, on average, improves performance by 11% compared to the state-of-the-art technique that manages cache partitioning and prefetching and by 50% compared to the baseline without cache partitioning, bandwidth partitioning and prefetch throttling.

The current trend for domain-specific architectures (DSAs) has led to renewed interest in research test chips to demonstrate new specialized hardware. Tape-outs also offer huge pedagogical value garnered from real hands-on exposure to the whole system stack. However, successful tape-outs demand hard-earned experience, and the design process is time consuming and fraught with challenges. Therefore, custom chips have remained the preserve of a small number of research groups, typically focused on circuit design research. This paper describes the CHIPKIT framework. We describe a reusable SoC subsystem which provides basic IO, an on-chip programmable host, memory and peripherals. This subsystem can be readily extended with new IP blocks to generate custom test chips. We also present an agile RTL development flow, including a code generation tool calledVGEN. Finally, we outline best practices for full-chip validation across the entire design cycle.