SiPearl unveils key features of the Rhea 1 to accelerate HPC and AI inference

1 min read

SiPearl, the company that’s building the high-performance low-power European microprocessor for HPC and AI inference has announced the main features of its first-generation microprocessor.

Credit: SiPearl

Designated Rhea1 the device will, according to the company, fulfill the mission entrusted by EuroHPC JU and the European Processor Initiative consortium: to bring dedicated high-performance microprocessor technologies back to Europe.”

Designed with high-performance energy-efficient arm Neoverse V1 platform, Rhea1 will include in a single package:

  • 80 arm Neoverse V1 cores ensuring high compute performance and efficient performance per watt. Each core includes 2 Scalable Vector Extension (SVE) of 256 bits each, enabling fast vector computations while optimising area and energy use;
  • Built-in High Bandwidth Memory, with 4 stacks of HBM, to provide a balanced solution ideal for HPC, big data and AI Inference applications, which are often memory bandwidth bound;
  • 4 DDR5 interfaces supporting 2 DIMMs Per Channel (2DPC);
  • 104 lanes of PCIe Gen5 interface: up to 6 x16 lanes + 2 x4 lanes;
  • High-performance arm Neoverse CMN-700 Coherent Mesh Network on Chip (NoC) to interconnect compute and I/O elements;
  • Support for Flat or Quadrant mode.

Rhea1 will, according SiPearl, be supported by a wide range of compilers, library and tools, from traditional programming languages such as C/C++, GO and RUST to modern AI frameworks such as TensorFlow or PyTorch.

Samples are scheduled for 2025.

Rhea1 is intended for traditional HPC workloads - its initial target market - and also to AI inference workloads.

The company employs more than 190 people in: France (Maisons-Laffitte, Grenoble, Massy, Sophia Antipolis), Germany (Duisburg), Italy (Bologna) and Spain (Barcelona)