Abstract We present a novel deep neural network (DNN) training scheme and resistive RAM (RRAM) in-memory computing (IMC) hardware evaluation towards achieving high accuracy against RRAM device/array variations and enhanced robustness against adversarial input attacks. We present improved IMC inference accuracy results evaluated on state-of-the-art DNNs including ResNet-18, AlexNet, and VGG with binary, 2-bit, and 4-bit activation/weight precision for the CIFAR-10 dataset. These DNNs are evaluated with measured noise data obtained from three different RRAM-based IMC prototype chips. Across these various DNNs and IMC chip measurements, we show that our proposed hardware noise-aware DNN training consistently improves DNN inference accuracy for actual IMC hardware, up to 8% accuracy improvement for the CIFAR-10 dataset. We also analyze the impact of our proposed noise injection scheme on the adversarial robustness of ResNet-18 DNNs with 1-bit, 2-bit, and 4-bit activation/weight precision. Our results show up to 6% improvement in the robustness to black-box adversarial input attacks.
more »
« less
This content will become publicly available on September 1, 2024
FP-IMC: A 28nm All-Digital Configurable Floating-Point In-Memory Computing Macro
In-memory computing (IMC) provides energy- efficient solutions to deep neural networks (DNN). Most IMC de- signs for DNNs employ fixed-point precisions. However, floating- point precision is still required for DNN training and complex inference models to maintain high accuracy. There have not been float-point precision based IMC works in the literature where the float-point computation is immersed into the weight memory storage. In this work, we propose a novel floating-point precision IMC macro with a configurable architecture that supports both normal 8-bit floating point (FP8) and 8-bit block floating point (BF8) with a shared exponent. The proposed FP-IMC macro implemented in 28nm CMOS demonstrates 12.1 TOPS/W for FP8 precision and 66.6 TOPS/W for BF8 precision, improving energy-efficiency beyond the state-of-the-art FP IMC macros.
more »
« less
- NSF-PAR ID:
- 10462009
- Date Published:
- Journal Name:
- Proceedings of ESSCIRC
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Mainstream math libraries for floating point (FP) do not produce correctly rounded results for all inputs. In contrast, CR-LIBM and RLIBM provide correctly rounded implementations for a specific FP representation with one rounding mode. Using such libraries for a representation with a new rounding mode or with different precision will result in wrong results due to double rounding. This paper proposes a novel method to generate a single polynomial approximation that produces correctly rounded results for all inputs for multiple rounding modes and multiple precision configurations. To generate a correctly rounded library for n -bits, our key idea is to generate a polynomial approximation for a representation with n +2-bits using the round-to-odd mode. We prove that the resulting polynomial approximation will produce correctly rounded results for all five rounding modes in the standard and for multiple representations with k -bits such that | E | +1 < k ≤ n , where | E | is the number of exponent bits in the representation. Similar to our prior work in the RLIBM project, we approximate the correctly rounded result when we generate the library with n +2-bits using the round-to-odd mode. We also generate polynomial approximations by structuring it as a linear programming problem but propose enhancements to polynomial generation to handle the round-to-odd mode. Our prototype is the first 32-bit float library that produces correctly rounded results with all rounding modes in the IEEE standard for all inputs with a single polynomial approximation. It also produces correctly rounded results for any FP configuration ranging from 10-bits to 32-bits while also being faster than mainstream libraries.more » « less
-
In-memory-computing (IMC) SRAM architecture has gained significant attention as it achieves high energy efficiency for computing a convolutional neural network (CNN) model [1]. Recent works investigated the use of analog-mixed-signal (AMS) hardware for high area and energy efficiency [2], [3]. However, AMS hardware output is well known to be susceptible to process, voltage, and temperature (PVT) variations, limiting the computing precision and ultimately the inference accuracy of a CNN. We reconfirmed, through the simulation of a capacitor-based IMC SRAM macro that computes a 256D binary dot product, that the AMS computing hardware has a significant root-mean-square error (RMSE) of 22.5% across the worst-case voltage, temperature (Fig. 16.1.1 top left) and 3-sigma process variations (Fig. 16.1.1 top right). On the other hand, we can implement an IMC SRAM macro using robust digital logic [4], which can virtually eliminate the variability issue (Fig. 16.1.1 top). However, digital circuits require more devices than AMS counterparts (e.g., 28 transistors for a mirror full adder [FA]). As a result, a recent digital IMC SRAM shows a lower area efficiency of 6368F2/b (22nm, 4b/4b weight/activation) [5] than the AMS counterpart (1170F2/b, 65nm, 1b/1b) [3]. In light of this, we aim to adopt approximate arithmetic hardware to improve area and power efficiency and present two digital IMC macros (DIMC) with different levels of approximation (Fig. 16.1.1 bottom left). Also, we propose an approximation-aware training algorithm and a number format to minimize inference accuracy degradation induced by approximate hardware (Fig. 16.1.1 bottom right). We prototyped a 28nm test chip: for a 1b/1b CNN model for CIFAR-10 and across 0.5-to-1.1V supply, the DIMC with double-approximate hardware (DIMC-D) achieves 2569F2/b, 932-2219TOPS/W, 475-20032GOPS, and 86.96% accuracy, while for a 4b/1b CNN model, the DIMC with the single-approximate hardware (DIMC-S) achieves 3814F2/b, 458-990TOPS/Wmore » « less
-
Obeid, I. ; Selesnik, I. ; Picone, J. (Ed.)The Neuronix high-performance computing cluster allows us to conduct extensive machine learning experiments on big data [1]. This heterogeneous cluster uses innovative scheduling technology, Slurm [2], that manages a network of CPUs and graphics processing units (GPUs). The GPU farm consists of a variety of processors ranging from low-end consumer grade devices such as the Nvidia GTX 970 to higher-end devices such as the GeForce RTX 2080. These GPUs are essential to our research since they allow extremely compute-intensive deep learning tasks to be executed on massive data resources such as the TUH EEG Corpus [2]. We use TensorFlow [3] as the core machine learning library for our deep learning systems, and routinely employ multiple GPUs to accelerate the training process. Reproducible results are essential to machine learning research. Reproducibility in this context means the ability to replicate an existing experiment – performance metrics such as error rates should be identical and floating-point calculations should match closely. Three examples of ways we typically expect an experiment to be replicable are: (1) The same job run on the same processor should produce the same results each time it is run. (2) A job run on a CPU and GPU should produce identical results. (3) A job should produce comparable results if the data is presented in a different order. System optimization requires an ability to directly compare error rates for algorithms evaluated under comparable operating conditions. However, it is a difficult task to exactly reproduce the results for large, complex deep learning systems that often require more than a trillion calculations per experiment [5]. This is a fairly well-known issue and one we will explore in this poster. Researchers must be able to replicate results on a specific data set to establish the integrity of an implementation. They can then use that implementation as a baseline for comparison purposes. A lack of reproducibility makes it very difficult to debug algorithms and validate changes to the system. Equally important, since many results in deep learning research are dependent on the order in which the system is exposed to the data, the specific processors used, and even the order in which those processors are accessed, it becomes a challenging problem to compare two algorithms since each system must be individually optimized for a specific data set or processor. This is extremely time-consuming for algorithm research in which a single run often taxes a computing environment to its limits. Well-known techniques such as cross-validation [5,6] can be used to mitigate these effects, but this is also computationally expensive. These issues are further compounded by the fact that most deep learning algorithms are susceptible to the way computational noise propagates through the system. GPUs are particularly notorious for this because, in a clustered environment, it becomes more difficult to control which processors are used at various points in time. Another equally frustrating issue is that upgrades to the deep learning package, such as the transition from TensorFlow v1.9 to v1.13, can also result in large fluctuations in error rates when re-running the same experiment. Since TensorFlow is constantly updating functions to support GPU use, maintaining an historical archive of experimental results that can be used to calibrate algorithm research is quite a challenge. This makes it very difficult to optimize the system or select the best configurations. The overall impact of all of these issues described above is significant as error rates can fluctuate by as much as 25% due to these types of computational issues. Cross-validation is one technique used to mitigate this, but that is expensive since you need to do multiple runs over the data, which further taxes a computing infrastructure already running at max capacity. GPUs are preferred when training a large network since these systems train at least two orders of magnitude faster than CPUs [7]. Large-scale experiments are simply not feasible without using GPUs. However, there is a tradeoff to gain this performance. Since all our GPUs use the NVIDIA CUDA® Deep Neural Network library (cuDNN) [8], a GPU-accelerated library of primitives for deep neural networks, it adds an element of randomness into the experiment. When a GPU is used to train a network in TensorFlow, it automatically searches for a cuDNN implementation. NVIDIA’s cuDNN implementation provides algorithms that increase the performance and help the model train quicker, but they are non-deterministic algorithms [9,10]. Since our networks have many complex layers, there is no easy way to avoid this randomness. Instead of comparing each epoch, we compare the average performance of the experiment because it gives us a hint of how our model is performing per experiment, and if the changes we make are efficient. In this poster, we will discuss a variety of issues related to reproducibility and introduce ways we mitigate these effects. For example, TensorFlow uses a random number generator (RNG) which is not seeded by default. TensorFlow determines the initialization point and how certain functions execute using the RNG. The solution for this is seeding all the necessary components before training the model. This forces TensorFlow to use the same initialization point and sets how certain layers work (e.g., dropout layers). However, seeding all the RNGs will not guarantee a controlled experiment. Other variables can affect the outcome of the experiment such as training using GPUs, allowing multi-threading on CPUs, using certain layers, etc. To mitigate our problems with reproducibility, we first make sure that the data is processed in the same order during training. Therefore, we save the data from the last experiment and to make sure the newer experiment follows the same order. If we allow the data to be shuffled, it can affect the performance due to how the model was exposed to the data. We also specify the float data type to be 32-bit since Python defaults to 64-bit. We try to avoid using 64-bit precision because the numbers produced by a GPU can vary significantly depending on the GPU architecture [11-13]. Controlling precision somewhat reduces differences due to computational noise even though technically it increases the amount of computational noise. We are currently developing more advanced techniques for preserving the efficiency of our training process while also maintaining the ability to reproduce models. In our poster presentation we will demonstrate these issues using some novel visualization tools, present several examples of the extent to which these issues influence research results on electroencephalography (EEG) and digital pathology experiments and introduce new ways to manage such computational issues.more » « less
-
Vector-matrix multiplication (VMM) is a core operation in many signal and data processing algorithms. Previous work showed that analog multipliers based on nonvolatile memories have superior energy efficiency as compared to digital counterparts at low-to-medium computing precision. In this paper, we propose extremely energy efficient analog mode VMM circuit with digital input/output interface and configurable precision. Similar to some previous work, the computation is performed by gate-coupled circuit utilizing embedded floating gate (FG) memories. The main novelty of our approach is an ultra-low power sensing circuitry, which is designed based on translinear Gilbert cell in topological combination with a floating resistor and a low-gain amplifier. Additionally, the digital-to-analog input conversion is merged with VMM, while current-mode algorithmic analog-to-digital circuit is employed at the circuit backend. Such implementations of conversion and sensing allow for circuit operation entirely in a current domain, resulting in high performance and energy efficiency. For example, post-layout simulation results for 400×400 5-bit VMM circuit designed in 55 nm process with embedded NOR flash memory, show up to 400 MHz operation, 1.68 POps/J energy efficiency, and 39.45 TOps/mm2 computing throughput. Moreover, the circuit is robust against process-voltage-temperature variations, in part due to inclusion of additional FG cells that are utilized for offset compensation.more » « less
