Ara: A 1 GHz+ Scalable and Energy-Efficient RISC-V Vector Processor with Multi-Precision Floating Point Support in 22 nm FD-SOI

Matheus A. Cavalcante, Fabian Schuiki, Florian Zaruba, Michael Schaffner, Luca Benini

Abstract

In this paper, we present Ara, a 64-bit vector processor based on the version 0.5 draft of RISC-V's vector extension, implemented in GlobalFoundries 22FDX FD-SOI technology. Ara's microarchitecture is scalable, as it is composed of a set of identical lanes, each containing part of the processor's vector register file and functional units. It achieves up to 97% FPU utilization when running a 256 x 256 double precision matrix multiplication on sixteen lanes. Ara runs at more than 1 GHz in the typical corner (TT/0.80V/25 °C) achieving a performance up to 33 DP-GFLOPS. In terms of energy efficiency, Ara achieves up to 41 DP-GFLOPS/W under the same conditions, which is slightly superior to similar vector processors found in literature. An analysis on several vectorizable linear algebra computation kernels for a range of different matrix and vector sizes gives insight into performance limitations and bottlenecks for vector processors and outlines directions to maintain high energy efficiency even for small matrix sizes where the vector architecture achieves suboptimal utilization of the available FPUs.

Introduction

As Dennard scaling has reached its physical limits, increasing performance by simply raising clock frequencies is no longer viable. Modern computer architects face the persistent Von Neumann Bottleneck, where the energy and time spent fetching and decoding instructions often dwarf the actual computation. While multi-core designs and GPUs (SIMT) have helped mitigate these issues, they often carry significant overhead for regular, data-parallel tasks like linear algebra or machine learning.

This paper revisits vector processing—an architecture style famous in 20th-century supercomputers—leveraging the modern, open-source RISC-V Vector Extension (RVV) to create a scalable, energy-efficient accelerator. Ara is designed to minimize instruction overhead by amortizing the cost of a single instruction over hundreds of operations, providing a high-performance alternative for data-parallel workloads.

System Architecture

Ara is implemented as a 64-bit vector unit that acts as a coprocessor to Ariane, an open-source RISC-V scalar core (RV64GC). The system is designed for modularity and scalability, allowing for different configurations depending on the target performance and area constraints.

System architecture of Ara integrated with the Ariane scalar core.

Lane-Based Scalability

The core of Ara's microarchitecture is its lane-based design. Ara is composed of $N$ identical lanes (tested at configurations of 2, 4, 8, and 16). Each lane contains a slice of the Vector Register File (VRF) and its own dedicated functional units, including a Floating-Point Unit (FPU), an Arithmetic Logic Unit (ALU), and a Multiplier. This modularity ensures that the design can be easily scaled to meet specific throughput requirements.

Register Management and Banking

To maintain high throughput, the functional units must be constantly fed with data. To prevent multiple units from stalling while attempting to access the same memory bank, Ara implements a "Barber’s Pole" banking scheme. This involves a shifted register-to-bank mapping that ensures elements of different vector registers are distributed across the available SRAM banks, minimizing access conflicts.

The 'Barber's Pole' banking scheme for vector registers.

Evaluation Methodology

The processor was evaluated using three representative kernels that span the "Roofline" model, ranging from memory-bound to compute-bound operations:

  1. MATMUL: $n \times n$ double-precision matrix multiplication (Compute-bound).
  2. DAXPY: A standard vector operation ($Y = \alpha X + Y$) used to test memory-bound performance.
  3. DCONV: A tensor convolution kernel based on the first layer of GoogLeNet, representing extremely compute-bound deep learning workloads.

These kernels were tested across various matrix and vector sizes to identify performance bottlenecks and efficiency limits.

Key Results and Performance Analysis

Ara demonstrates exceptional performance and efficiency in 22 nm FD-SOI technology:

  • Clock Speed: The design achieved over 1 GHz in typical operating conditions (TT/0.80V/25 °C).
  • Peak Throughput: The 16-lane configuration reached up to 33 DP-GFLOPS.
  • Energy Efficiency: Ara achieved 41 DP-GFLOPS/W, making it approximately 2.5x more efficient than the Ariane scalar core running the same tasks.
  • Utilization: For large matrix multiplications, Ara achieved a staggering 97% FPU utilization, indicating that the architecture effectively hides memory latency and instruction overhead.

Discussion and Future Directions

One of the primary insights from this research is the identification of the Scalar Bottleneck. For small vector lengths, performance is limited not by the vector unit's capacity, but by the scalar core's ability to issue instructions fast enough. If the scalar core cannot "feed" Ara at a sufficient rate, the vector lanes remain idle, reducing overall efficiency. Future work may explore the use of superscalar or VLIW scalar cores to improve this issue rate.

Additionally, while Ara is highly efficient, it currently lacks native hardware support for vector reduction instructions (e.g., summing all elements in a vector), which must be emulated. Implementing lane-level power gating and native reduction support are logical next steps to further enhance the processor's versatility and energy profile in edge computing and high-performance environments.

Summary updated

Abstract

In this paper, we present Ara, a 64-bit vector processor based on the version 0.5 draft of RISC-V's vector extension, implemented in GlobalFoundries 22FDX FD-SOI technology. Ara's microarchitecture is scalable, as it is composed of a set of identical lanes, each containing part of the processor's vector register file and functional units. It achieves up to 97% FPU utilization when running a 256 x 256 double precision matrix multiplication on sixteen lanes. Ara runs at more than 1 GHz in the typical corner (TT/0.80V/25 oC) achieving a performance up to 33 DP-GFLOPS. In terms of energy efficiency, Ara achieves up to 41 DP-GFLOPS/W under the same conditions, which is slightly superior to similar vector processors found in literature. An analysis on several vectorizable linear algebra computation kernels for a range of different matrix and vector sizes gives insight into performance limitations and bottlenecks for vector processors and outlines directions to maintain high energy efficiency even for small matrix sizes where the vector architecture achieves suboptimal utilization of the available FPUs.
BibTeX
@article{cavalcante2020ara,
  author = {Matheus A. Cavalcante and Fabian Schuiki and Florian Zaruba and Michael Schaffner and Luca Benini},
  title = {{Ara: A 1 GHz+ Scalable and Energy-Efficient RISC-V Vector Processor with Multi-Precision Floating Point Support in 22 nm FD-SOI}},
  journal = {IEEE Trans. Very Large Scale Integr. Syst.},
  year = {2020},
  doi = {10.1109/tvlsi.2019.2950087},
}
Figure