An Accelerator for High Efficient Vision Processing

In recent years, neural network accelerators have been shown to achieve both high energy efficiency and high performance for a broad application scope within the important category of recognition and mining applications. Still, both the energy efficiency and performance of such accelerators remain limited by memory accesses. In this paper, we focus on image applications, arguably the most important category among recognition and mining applications. The neural networks which are state-of-the-art for these applications are convolutional neural networks (CNNs), and they have an important property: weights are shared among many neurons, considerably reducing the neural network memory footprint. This property allows to entirely map a CNN within an SRAM, eliminating all DRAM accesses for weights. By further hoisting this accelerator next to the image sensor, it is possible to eliminate all remaining DRAM accesses, i.e., for inputs and outputs. In this paper, we propose such a CNN accelerator, placed next to a CMOS or CCD sensor. The absence of DRAM accesses combined with a careful exploitation of the specific data access patterns within CNNs allows us to design an accelerator which is highly energy-efficient. We present a single-core implementation down to the layout at 65 nm, with a modest footprint of 5.94mm(2) and consuming only 336mm(2), but still about 30x faster than high-end GPUs. For visual processing with higher resolution and frame-rate requirements, we further present a multicore implementation with elevated performance.

Miss-Optimized Memory Systems: Turning Thousands of Outstanding Misses into Reuse Opportunities

Mikhail Asiatici

Even if Dennard scaling came to an end fifteen years ago, Moore's law kept fueling an exponential growth in compute performance through increased parallelization. However, the performance of memory and, in particular, Dynamic Random Access Memory (DRAM), has been increasing at a slower pace for decades, making memory system optimization increasingly crucial. Conventional solutions mitigate the issue by shifting as many memory accesses as possible from off-chip DRAM to on-chip Static RAM (SRAM) memory, which has higher performance but lower capacity. This is achieved by relying on spatial and temporal locality or on precise compile-time information about the access pattern. However, when the access pattern is irregular and data-dependent, these solutions are ineffective and the processing-memory gap grows even wider asDRAMs themselves are optimized for sequential accesses.In this thesis, we present a novel memory system for throughput-oriented compute engines that perform irregular read accesses to DRAM. When accesses are irregular, we acknowledge that obtaining a reasonable benefit from on-chip memory may be unrealistic; therefore, we focus on minimizing stalls and reusing each memory response to serve as many misses as possible without relying on long-term data storage. This is the same insight behind nonblocking caches but on a vastly larger scale in terms of outstanding misses, which greatly increases the opportunities for data reuse when accelerators emit a large number of outstanding reads.Because we optimize miss handling rather than increasing hit rate, we call our architecture miss-optimized memory system (MOMS). We first focus on the microarchitectural level to show how a MOMS can support three orders of magnitude more outstanding misses than a traditional nonblocking cache in a way that can be efficiently implemented on Field-Programmable Gate Arrays (FPGAs). Once we maximize the reuse of each individual word returned by the DRAM, we introduce two techniques to increase the DRAM throughput. When the DRAM controller is optimized for burst requests, we group incoming requests over multiple words that are requested as a burst. Conversely, when the DRAM controller handles single requests efficiently, our MOMS reorders requests by DRAM bank and row on a much larger scale than general-purpose DRAM controllers. We then discuss techniques to use efficiently the vast amount of resources provided by multi-die FPGAs and introduce two-level architectures which balance reuse maximization and contention of shared hardware. Finally, we develop a graph processing accelerator backed by a MOMS. On three algorithms on graphs with billions of edges and up to a hundred million nodes, our accelerator outperforms the state-of-the-art on FPGAs and achieves higher performance per watt and unit bandwidth than the state-of-the-art on CPUs and GPUs.Memory systems designers face increasing pressure to keep up with the compute engines performance. Our results suggest that miss-optimized memory systems can help to reduce the memory-processing gap where it is largest, that is, when accesses to memory are irregular and difficult to serve from local buffers.

EPFL2021

Adding limited reconfigurability to superscalar processors

Marc Epalza

For the last thirty years, electronics, at first built with discrete components, and then as Integrated Circuits (IC), have brought diverse and lasting improvements to our quality of life. Examples might include digital calculators, automotive and airplane control assistance, almost all electrical household appliances, and the almost ubiquitous Personal Computer. Application-Specific Integrated Circuits (ASICs) were traditionally used for their high performance and low manufacturing cost, and were designed specifically for a single application with large volumes. But as lower product lifetimes and the pressures of fast marketing increased, ASICs' high design cost pushed for their replacement by Microprocessors. These processors, capable of implementing any functionality through a change in software, are thus often called General Purpose Processors. General purpose processors are used for everyday computing tasks, and found in all personal computers. They are also often used as building blocks for scientific supercomputers. Superscalar processors such as these require ever more processing power to run complex simulations, video games or versatile telecoms services. In the case of embedded applications, e.g. for portable devices, both performance and power consumption must be taken into account. In a bid to adapt a processor to some extent to select applications, fully reconfigurable logic can greatly improve the performance of a processor, since it is shaped for the best possible execution with the available resources. However, as reconfigurable logic is far slower than custom logic, this gain is possible only for some specific applications with large parallelism, after a detailed study of the algorithm. Even though this process can be automated, it still requires large computing resources, and cannot be performed at run time. To reduce the loss in speed compared to custom logic, it is possible to limit the reconfigurability to increase the breadth of applications where performance can be improved. However, as the application space increases, a careful analysis and design of the reconfigurability is required to minimize the speed loss, notably when dynamic reconfiguration is considered. As a case study, we analyze the feasibility of adding limited reconfigurability to the Floating Point Units (FPUs) of a general purpose processor. These rather large units execute all floating point operations, and may also be used for integer multiplication. If an application contains few or infrequent instructions that must be executed by the FPU, this idle hardware only increases power consumption without enhancing performance. This is often the case in non-scientific applications and even many recent and detailed video games which make heavy use of hardware display accelerators for 3D graphics. In a fast multiplier such as can be found in the FPU of a high performance processor, the logic to perform multiplication is a large tree of compressors to add all the partial products together. It is possible to add logic to allow the reconfiguration of part of this tree as several extra Arithmetic and Logic Units (ALU). This requires a detailed timing analysis for both the reconfigurable FPU and the extra ALUs, taking into account effects such as added wires and longer critical paths. Finally, the algorithm to decide when and how to reconfigure must be studied, in terms of eciency and complexity. The results of adding this limited reconfigurability to a mainstream superscalar processor over a large set of compute intensive benchmarks show gains of up to 56% in the best case, with an average gain of 11%. The application to an idealized huge top processor still shows slightly positive average gains, as the limits of available parallelism are reached, bounded by both the application and many of the characteristics of the processor. In all cases, binary compatibility is maintained, allowing the re-use of all existing software. We show that adding limited reconfigurability to a general purpose superscalar processor can produce interesting gains over a wide range of applications while maintaining binary compatibility, and without large modifications to the original design. Limited reconfigurability is worthwhile as it increases the design space, allowing gains to apply to a larger set of applications. These gains are achieved through careful study and optimization of the reconfigurable logic and the decision algorithm.

EPFL2004

Memory Systems and Interconnects for Scale-Out Servers

Stavros Volos

The information revolution of the last decade has been fueled by the digitization of almost all human activities through a wide range of Internet services. The backbone of this information age are scale-out datacenters that need to collect, store, and process massive amounts of data. These datacenters distribute vast datasets across a large number of servers, typically into memory-resident shards so as to maintain strict quality-of-service guarantees. While data is driving the skyrocketing demands for scale-out servers, processor and memory manufacturers have reached fundamental efficiency limits, no longer able to increase server energy efficiency at a sufficient pace. As a result, energy has emerged as the main obstacle to the scalability of information technology (IT) with huge economic implications. Delivering sustainable IT calls for a paradigm shift in computer system design. As memory has taken a central role in IT infrastructure, memory-centric architectures are required to fully utilize the IT's costly memory investment. In response, processor architects are resorting to manycore architectures to leverage the abundant request-level parallelism found in data-centric applications. Manycore processors fully utilize available memory resources, thereby increasing IT efficiency by almost an order of magnitude. Because manycore server chips execute a large number of concurrent requests, they exhibit high incidence of accesses to the last-level-cache for fetching instructions (due to large instruction footprints), and off-chip memory (due to lack of temporal reuse in on-chip caches) for accessing dataset objects. As a result, on-chip interconnects and the memory system are emerging as major performance and energy-efficiency bottlenecks in servers. This thesis seeks to architect on-chip interconnects and memory systems that are tuned for the requirements of memory-centric scale-out servers. By studying a wide range of data-centric applications, we uncover application phenomena common in data-centric applications, and examine their implications on on-chip network and off-chip memory traffic. Finally, we propose specialized on-chip interconnects and memory systems that leverage common traffic characteristics, thereby improving server throughput and energy efficiency.

EPFL2015