Search

Recurrent Neural Networks (RNNs) are a key technology for emerging applications such as automatic speech recognition, machine translation or image description. Long Short Term Memory (LSTM) networks are the most successful RNN implementation, as they can learn long term dependencies to achieve high accuracy. Unfortunately, the recurrent nature of LSTM networks significantly constrains the amount of parallelism and, hence, multicore CPUs and many-core GPUs exhibit poor efficiency for RNN inference. In this paper, we present E-PUR, an energy-efficient processing unit tailored to the requirements of LSTM computation. The main goal of E-PUR is to support large recurrent neural networks for low-power mobile devices. E-PUR provides an efficient hardware implementation of LSTM networks that is flexible to support diverse applications. One of its main novelties is a technique that we call Maximizing Weight Locality (MWL), which improves the temporal locality of the memory accesses for fetching the synaptic weights, reducing the memory requirements by a large extent. Our experimental results show that E-PUR achieves real-time performance for different LSTM networks, while reducing energy consumption by orders of magnitude with respect to general-purpose processors and GPUs, and it requires a very small chip area. Compared to a modern mobile SoC, an NVIDIA Tegra X1, E-PUR provides an average energy reduction of 88x.

Evaluation techniques in microprocessor design are mostly based on simulating selected application samples using a cycle-accurate simulator. In order to achieve accurate results, microarchitectural structures are warmed-up for a few million instructions prior to statistics collection. Unfortunately, this strategy cannot be applied to HW/SW co-designed processors, in which a Transparent Optimization software Layer (TOL) translates and optimizes code on-the-fly from a guest ISA to an internal host custom microarchitecture. We show that the warm-up period in this case needs to be 3-4 orders of magnitude longer than what is needed for traditional microprocessor designs because the TOL state needs to be warmed-up as well. In this paper, we propose a novel simulation technique for HW/SW co-designed processors based on adapting the optimization promotion thresholds using high level application statistics in order to find the best trade-off between accuracy and simulation cost. In particular, the proposed technique reduces the simulation cost by 65X with an average error of just 0.75%. Furthermore, as opposed to other alternatives, the proposed technique satisfies the additional requirement of allowing evaluation using different TOL and microarchitectural configurations.

The heavily-threaded data processing demands of streaming multiprocessors (SM) in a GPGPU require a large register file (RF). The fast increasing size of the RF makes the area cost and power consumption unaffordable for traditional SRAM designs in the future technologies. In this paper, we propose to use embedded-DRAM (eDRAM) as an alternative in future GPGPUs. Compared with SRAM, eDRAM provides higher density and lower leakage power. However, the limited data retention time in eDRAM poses new challenges. Periodic refresh operations are needed to maintain data integrity. This is exacerbated with the scaling of eDRAM density, process variations and temperature. Unlike conventional CPUs which make use of multi-ported RF, most of the RFs in modern GPGPU are heavily banked but not multi-ported to reduce the hardware cost. This provides a unique opportunity to hide the refresh overhead. We propose two different eDRAM implementations based on 3T1D and 1T1C memory cells. To mitigate the impact of periodic refresh, we propose two novel refresh solutions using bank bubble and bank walk-through. Plus, for the 1T1C RF, we design an interleaved bank organization together with an intelligent warp scheduling strategy to reduce the impact of the destructive reads. The analysis shows that our schemes present better energy efficiency, scalability and variation tolerance than traditional SRAM-based designs.

In this paper we present TEAPOT, a full system GPU simulator, whose goal is to allow the evaluation of the GPUs that reside in mobile phones and tablets. To this extent, it has a cycle accurate GPU model for evaluating performance, power models for the GPU, the memory subsystem and for OLED screens, and image quality metrics. Unlike prior GPU simulators, TEAPOT supports the OpenGL ES 1.1/2.0 API, so that it can simulate all commercial graphical applications available for Android systems. To illustrate potential uses of this simulating infrastructure, we perform two case studies. We first turn our attention to evaluating the impact of the OS when simulating graphical applications. We show that the overall GPU power/performance is greatly aff ected by common OS tasks, such as image composition, and argue that application level simulation is not sufficient to understand the overall GPU behavior. We then utilize the capabilities of TEAPOT to perform studies that trade image quality for energy. We demonstrate that by allowing for small distortions in the overall image quality, a signifi cant amount of energy can be saved.

Dynamic Binary Translators (DBT) and Dynamic Binary Opti- mization (DBO) by software are used widely for several reasons including performance, design simplification and virtualization. However, the software layer in such systems introduces non- negligible overheads which affect performance and user experi- ence. Hence, reducing DBT/DBO overheads is of paramount im- portance. In addition, reduced overheads have interesting collateral effects in the rest of the software layer, such as allowing optimiza- tions to be applied earlier. A cost-effective solution to this problem is to provide hardware support to speed up the primitives of the software layer, paying special attention to automate DBT/DBO mechanisms and leave the heuristics to the software, which is more flexible. In this work, we have characterized the overheads of a DBO sys- tem using DynamoRIO implementing several basic optimizations. We have seen that the computation of the Data Dependence Graph (DDG) accounts for 5%-10% of the execution time. For this rea- son, we propose to add hardware support for this task in the form of a new functional unit, called DDGacc, which is integrated in a conventional pipeline processor and is operated through new ISA instructions. Our evaluation shows that DDGacc reduces the cost of computing the DDG by 32x, which reduces overall execution time by 5%-10% on average and up to 18% for applications where the DBO optimizes large code footprints.

SRAM and DRAM cells have been the predominant technologies used to implement memory cells in computer systems, each one having its advantages and shortcomings. SRAM cells are faster and require no refresh since reads are not destructive. In contrast, DRAM cells provide higher density and minimal leakage energy since there are no paths within the cell from Vdd to ground. Recently, DRAM cells have been embedded in logic-based technology, thus overcoming the speed limit of typical DRAM cells. In this paper we propose an n-bit macrocell that implements one static cell, and n-1 dynamic cells. This cell is aimed at being used in an n-way set-associative first-level data cache. Our study shows that in a four-way set-associative cache with this macrocell compared to an SRAM based with the same capacity, leakage is reduced by about 75% and area more than half with a minimal impact on performance. Architectural mechanisms have also been devised to avoid refresh logic. Experimental results show that no performance is lost when the retention time is larger than 50K processor cycles. In addition, the proposed delayed writeback policy that avoids refreshing performs a similar amount of writebacks than a conventional cache with the same organization, so no power wasting is incurred.