Ricerc@Sapienza

In this paper, an emulation environment for approximate memory architectures is presented. In the context of error tolerant applications, in which energy is saved at the expense of the occurrence of errors in data processing, approximate memories play a relevant part. Approximate memories are memories where read/write errors are allowed with controlled probability. In general these errors are the result of circuital or architectural techniques (i.e. voltage scaling, refresh rate reduction) introduced to save energy.

In this work we present AppropinQuo, a flexible and configurable emulator for embedded platforms with approximate memory. The emulator includes models of the effects of approximate memory circuits and architectures, that depend on the internal structure and organization of the cells. The ability to emulate a complete platform, including CPU, peripherals and hardware-software interactions, is particularly important since it allows to execute the application as on the real board, reproducing the effects of errors on output.

Improving power consumption and performance of error tolerant applications is the target of the design paradigm known as approximate computing. The memory subsystem is one of the units of a computational architecture where approximations can be introduced, leveraging on the resilience of an application to maintain an acceptable output quality even if its input data are subject to imprecision and errors.This paper proposes and implements the management, in the Linux kernel, of multiple approximate memory banks.

This paper presents the design and the implementation of a fully combinatorial floating point unit (FPU). The FPU can be reconfigured at implementation time in order to use an arbitrary number of bits for the mantissa and exponent, and it can be synthesized in order to support all IEEE-754 compliant FP formats but also non-standard FP formats, exploring the trade-off between precision (mantissa field), dynamic range (exponent field) and physical resources.

Approximate computing is a new paradigm for energy efficient design, based on the idea of designing digital systems that trade off computational accuracy for energy consumption. The paradigm can be applied to different units (i.e. internal units of the CPU, floating point coprocessors, memories). Considering the memory subsystem, approximate memories are physical memories where circuit-level or architecture-level techniques are implemented in order to reduce energy at the expense of errors occurring in bit cells.