In the context of highly data-centric applications, close reconciliation of computation and storage should significantly reduce the energy-consuming process of data movement. This paper proposes a Computational SRAM (CSRAM) combining In-and Near-Memory Computing (IMC/NMC) approaches to be used by a scalar processor as an energy-efficient vector processing unit. Parallel computing is thus performed on vectorized integer data on large words using usual logic and arithmetic operators. Furthermore, multiple rows can be advantageously activated simultaneously to increase this parallelism. The proposed C-SRAM is designed with a two-port pushed-rule foundry bitcell, available in most existing design platforms, and an adjustable form factor to facilitate physical implementation in a SoC. The 4kB C-SRAM testchip of 128-bit words manufactured in 22nm FD-SOI process technology displays a sub-array efficiency of 72% as well as an additional computing area of less than 5%. The measurements averaged on 10 dies at 0.85V and 1GHz demonstrate an energy efficiency per unit area of 35.6 and 1.48TOPS/W/mm2 for 8-bit additions and multiplications with 3ns and 24ns computing latency, respectively. Compared to a 128-bit SIMD processor architecture, up to 2x energy reduction and 1.8x speed-up gains are achievable for a representative set of highly data-centric application kernels.
In-memory computing (IMC) aims to solve the performance gap between CPU and memories introduced by the memory wall. However, it does not address the energy wall problem caused by data transfer over memory hierarchies. This paper proposes the data-locality management unit (DMU) to efficiently transfer data from a DRAM memory to a computational SRAM (C-SRAM) memory allowing IMC operations. The DMU is tightly coupled within the C-SRAM and allows one to align the data structure in order to perform effective in-memory computation. We propose a dedicated instruction set within the DMU to issue data transfers. The performance evaluation of a system integrating C-SRAM within the DMU compared to a reference scalar system architecture shows an increase from to in speed-up and from to in energy reduction, versus a system integrating C-SRAM without any transfer mechanism compared to a reference scalar system architecture.
30 March 2023 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
In-memory computing (IMC) aims at solving the performance gap between CPU and memories introduced by the memory wall. However, general-purpose IMC does not consider the optimization of data transfers for patterns, such as stencils and convolutions. This article proposes a new instruction set architecture (ISA) and a novel pattern encoding for IMC to transfer and organize data streams in order to perform efficiently computation. This instruction set is implemented on the data-locality management unit (DMU) as a subset of the computational SRAM (C-SRAM) ISA. A programming model to interact with the DMU at language level is also presented in this article. This DMU ISA is evaluated on six applications run on three different system nodes. These system nodes are based on existing RISC-V cores and range from embedded to high-performance computing domain. Experiments show on average a speed-up of $\times 8.81$ , an energy reduction factor of $\times 6.81$ , and an improvement of the number of operations per cycle of $\times 4.59$ , for the C-SRAM architecture integrating the proposed ISA of the DMU compared to a reference implementation on embedded systems. Results also show an improvement of the number of operations per cycle of $\times 2.99$ compared to a reference implementation on all system nodes.
This article describes a software environment called HybroGen, which helps to experiment binary code generation at run time. As computing architectures are getting more complex, the application performance is becoming data-dependent. The proposed experimental platform is helpful in programming applications that can be reconfigured at run time in order to be adapted for a new data environment. The HybroGen platform is adapted to heterogeneous architectures and can generate instructions for different targets. This platform allows to go farther than classical JIT compilation in many directions: the code generator is smaller by three orders of magnitude and faster by three orders of magnitude, compared to JIT (Just-In-Time) platforms, and allows making code transformation that is impossible in traditional compilation schemes, such as code generation for non von Neumann accelerators or dynamic code transformations for transprecision. The latter is illustrated in a code example: the square root with Newton’s algorithm. We also illustrate the proposed HybroGen platform with two other examples: a multiplication with a specialization on a value determined at run time, and a conversion of degrees Celsius to degrees Fahrenheit. This article presents a proof of concept of the proposed HybroGen platform in terms of its functionalities, and demonstrates the working status.
Adapting a source code to the specificity of its host hardware represents one way to implement software optimization. This allows to benefit from processors that are primarily designed to improve system performance. To reach such a software/hardware fitting without narrowing the scope of the optimization to few executions, one needs to have at his disposal relevant performance models of the considered hardware. This paper proposes a new method to optimize software kernels by considering their data-access mode. The proposed method permits to build a data-cache-miss model of a given application regarding its specific memory-access pattern. We apply our method in order to evaluate some custom implementations of matrix data layouts. To validate the functional correctness of the generated models, we propose a reference algorithm that simulates a kernel's exploration of its data. Experimental results show that the proposed data alignment permits to reduce the number of cache misses by a factor up to 50%, and to decrease the execution time by up to 30%. Finally, we show the necessity to integrate the impact of the Translation Lookaside Buffers (TLB) and the memory prefetcher within our performance models.