Internship - Digital circuit design for artificial intelligence - M/F

  • Artificial Intelligence & data intelligence,
  • Internship
  • 5 à 6 mois
  • CEA-List
  • Grenoble
  • Level 7
  • 2027-02-01
Apply

Context: As part of our ASIC design activities, we are developing an embedded computing platform based on a RISC-V processor. For near-sensor data processing, we use transformer-based artificial intelligence (AI) algorithms. These models significantly improve performance but require high computational power. To increase the platform's computing capacity while adhering to a strict power budget, integrating optimized hardware accelerators can be a solution, particularly for matrix-vector and matrix-matrix multiplication operations. Additionally, due to limited memory, it is necessary to compress AI models, for example by storing network weights in low-bit binary formats (less than one byte), which is below the granularity of a general-purpose processor. Hardware accelerators must therefore support these representations. Internship objective: The main objective of the internship is to integrate a reduced-precision matrix multiplication accelerator into the RTL (SystemVerilog) platform. In addition to hardware design, the internship will include demonstrating the software environment that enables controlling computations on the accelerator from the embedded processor. Based on an existing design, the accelerator can be optimized to support new binary formats, such as 4-bit floating-point (FP4). Work plan: v  Mastering the hardware architecture of the platform and the simulation environment (RTL level) v  Study of the compute accelerator and integration solutions within the platform v  Integration of the accelerator, implementation of the software flow for deploying computations on the accelerator, and functional validation v  Silicon projection (performance, power consumption, area) using a logic synthesis flow v  Performance estimation, energy efficiency assessment for target applications v  Evolution of the accelerator design to integrate new binary formats and revisiting previous steps

Le Laboratoire Intelligence Intégrée Multi-capteurs (LIIM) a pour mission d'étudier et de développer des algorithmes embarqués d'intelligence artificielle, de fusion de données et de perception de l'environnement pour des systèmes cyber-physiques multi-capteurs, de concevoir et 'implémenter des plateformes de démonstration logicielles et matérielles mettant en œuvre ces algorithmes avec l'introduction de technologies innovantes, notamment pour la réalisation de circuits intégrés spécifiques

Profil: Fin d'étude d'ingénieur ou Master 2 (Bac+5) Compétences en architecture de systèmes embarqués, conception de systèmes numériques, design RTL, logiciel embarqué Languages VHDL, SystemVerilog, C/C++, Python Connaissances sur l'intelligence artificielle, réseaux de neurones utiles

Bac+5 - Diplôme École d'ingénieurs

English Intermediate

en_USEN

Contact us

We will reply as soon as possible...