APROPOS at HiPEAC 2023

The 2023 edition of the HiPEAC conference on computer architecture, programming, compilers, and operating systems took place in Toulouse and comprised three keynotes from industry experts, several full paper sessions, and numerous topic-relevant workshops. Particularly, four of our ESRs had the opportunity to have their research presented and discuss early results or in-progress work to an interested audience and each other in Monday’s Workshop on Approximate Computing (WAPCO) session.

ESR 7 Sepide Saeedi was unfortunately unable to take part in the conference. More details about the HiPEAC conference are available at https://www.hipeac.net/2023/toulouse/#/

WAPCO presentations

For our ESRs, the WAPCO session was at the center of the conference. For this reason, we provide brief introductions to each their presentations below. Contact the ESRs for more information on each topic.

Automatic Generation of Input-Aware Approximate Arithmetic Circuits” presented by Ali Piri (ESR 6, École Centrale de Lyon, France)

Ali described how Approximate Computing (AxC) is applied systematically at various abstraction levels to reduce overheads and increase application performance but often without taking input distributions into account. Motivated by this, they proposed an input-aware approximate design approach that enables further approximation. Leveraging state-of-the-art approximation techniques and multi-objective optimization, this approach can further savings both in area and power consumption, illustrated by up to a 40% increase in an Artificial Intelligence (AI) application.

The library to generate input-aware approximate circuits is available at https://github.com/SalvatoreBarone/pyALS.

Fast Exploration of the Impact of Precision Reduction on Spiking Neural Networks” presented by Alessio Carpegna (in place of Sepide Saeedi, ESR 7, Politecnico di Torino, Italy)

The presenter covered how AxC techniques trade off computation accuracy for performance gains, and how applications intrinsically tolerant to some accuracy loss are most suitable as targets for such techniques. Spiking Neural Networks (SNNs) are one such application that is also highly relevant to Edge computing scenarios, requiring minimal area. They proposed using interval arithmetic concepts to model error propagation through an SNN and detect when the accuracy exceeds tolerable limits. This approach allows them to reduce approximation exploration time and reduce the size of network parameters further.

Toward Matrix Multiplication for Deep Learning Inference on the Xilinx Versal” presented by Jie Lei (ESR 9, Valencia Polytechnic University, Spain)

Jie demonstrated the optimization principles underlying the modern implementation of General Matrix Multiplication (GEMM) in conventional processor architectures. Later, they described how to achieve high performance for the type of operations that arise in Deep Learning (DL) on a hardware accelerator such as the AI Engine (AIE) tile embedded in Xilinx Versal platforms. Their experimental results with a prototype implementation of the GEMM kernel on a Xilinx Versal VCK190 show performance close to 86.7% of the theoretical peak of an AIE tile for 16-bit integer operands.

Verification of Approximate Hardware Designs” presented by Hans Jakob Damsgaard (ESR 10, Tampere University, Finland)

Hans Jakob was motivated by a lack of standardized verification flows for approximate hardware designs, despite many being evaluated by the same error metrics, and proposed an extension to an existing verification library, ChiselVerify, with support for just this. The extension hides sampling and metric computations behind a very simple Application Programmer Interface (API) that lets its users avoid unnecessary re-declarations of both. To demonstrate this functionality, they showcased a test bench for three approximate adders and reported error results for different configurations of them.

The library is available at https://github.com/chiselverify/chiselverify. Code to regenerate the paper’s results is available at https://github.com/hansemandse/wapco.

Acronyms:

  • AxC: Approximate Computing
  • AI: Artificial Intelligence
  • AIE: AI Engine
  • API: Application Programmer Interface
  • GEMM: General Matrix Multiplication
  • SNN: Spiking Neural Network

Sepide Saeedi

  • ESR 7
  • Politecnico di Torino - Arduino
More information

Jie Lei

  • ESR 9
  • Universitat Politecnica de Valencia - IBM
More information

Hans Jakob Damsgaard

  • ESR 10
  • Tampere University - ISW
More information

Ali Piri

  • ESR 6
  • Ecole Centrale de Lyon - Centro Regionale Information Communication Technology scrl
More information