HARDWARE – VEC

Vector Accelerator (VEC)

A many-core RISC-V vector chiplet for high-precision HPC and emerging applications that combine HPC and AI.

Lead partner

At a glance

Out-of-order RISC-V core

Superscalar core implementing the RISC-V Vector extension 1.0 (RVV 1.0).

Long-vector unit

Multi-lane vector processing unit with 16 kbit vector length.

Matrix extension for AI

Supports AI workloads on the same chip.

Many-core, multi-chiplet

Many cores per chiplet; several chiplets per package.

HBM memory

High-bandwidth memory for data-intensive workloads.

What we are building

The Vector Accelerator (VEC) is DARE’s RISC-V chiplet for high-precision HPC and emerging HPC-AI convergence applications. It is led by Openchip; Barcelona Supercomputing Center (BSC) leads the VEC pathfinding activities.

Each VEC core combines an out-of-order superscalar RISC-V core with a multi-lane long-vector unit and a matrix extension for AI. Many cores are combined in one chiplet, several chiplets are integrated with HBM memory in a System-in-Package, and the package is mounted on a PCIe card.

From VEC core to PCIe card. Chiplet, package and card are symbolic representations

Why it matters

High-precision HPC

Built for simulation workloads that need high numerical precision.

HPC meets AI

A matrix extension brings AI capability to the same vector chiplet.

Tested on real applications

IFS and NEMO (weather and climate, ECMWF), GROMACS (molecular dynamics) and Quantum ESPRESSO (quantum chemistry) are being optimised for VEC.