Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Tensor-Core Beamformer: A High-Speed Signal-Processing Library for Multidisciplinary Use

Leon Oostrum, Bram Veenboer, Ronald Rook, Michael Brown, Pieter Kruizinga, John W. Romein

TL;DR

The paper tackles the computational and energy challenges of real-time, large-scale beamforming across multiple domains. It introduces the Tensor-Core Beamformer (TCBF) and its domain-independent ccglib library, which maps beamforming to complex-valued GEMMs on GPU tensor cores with 16-bit and 1-bit precision, available on CUDA and HIP backends. Key contributions include a high-performance, energy-efficient GEMM library, an auto-tuning framework with thorough roofline analysis, and demonstrated applications in computational ultrasound and LOFAR radio astronomy that achieve substantial speedups and energy savings over prior implementations. The work significantly broadens practical beamforming capability on modern GPUs, enabling real-time 3D ultrasound feedback and more scalable radio-astronomy processing, while outlining avenues for future precision support and architectural optimizations.

Abstract

Beamforming is a well-known technique to combine signals from multiple sensors. It has a wide range of application domains. This paper introduces the Tensor-Core Beamformer: a generic, optimized beamformer library that harnesses the computational power of GPU tensor cores to accelerate beamforming computations. The library hides the complexity of tensor cores from the user, and supports 16-bit and 1-bit precision. An extensive performance evaluation on NVIDIA and AMD GPUs shows that the library outperforms traditional beamforming on regular GPU cores by a wide margin, at much higher energy efficiency. In the 16-bit mode, it achieves over 600 TeraOps/s on an AMD MI300X GPU, while approaching 1 TeraOp/J. In the 1-bit mode, it breaks the 3 PetaOps/s barrier and achieves over 10 TeraOps/J on an NVIDIA A100 GPU. The beamforming library can be easily integrated into existing pipelines. We demonstrate its use for medical ultrasound and radio-astronomical instruments.

The Tensor-Core Beamformer: A High-Speed Signal-Processing Library for Multidisciplinary Use

TL;DR

The paper tackles the computational and energy challenges of real-time, large-scale beamforming across multiple domains. It introduces the Tensor-Core Beamformer (TCBF) and its domain-independent ccglib library, which maps beamforming to complex-valued GEMMs on GPU tensor cores with 16-bit and 1-bit precision, available on CUDA and HIP backends. Key contributions include a high-performance, energy-efficient GEMM library, an auto-tuning framework with thorough roofline analysis, and demonstrated applications in computational ultrasound and LOFAR radio astronomy that achieve substantial speedups and energy savings over prior implementations. The work significantly broadens practical beamforming capability on modern GPUs, enabling real-time 3D ultrasound feedback and more scalable radio-astronomy processing, while outlining avenues for future precision support and architectural optimizations.

Abstract

Beamforming is a well-known technique to combine signals from multiple sensors. It has a wide range of application domains. This paper introduces the Tensor-Core Beamformer: a generic, optimized beamformer library that harnesses the computational power of GPU tensor cores to accelerate beamforming computations. The library hides the complexity of tensor cores from the user, and supports 16-bit and 1-bit precision. An extensive performance evaluation on NVIDIA and AMD GPUs shows that the library outperforms traditional beamforming on regular GPU cores by a wide margin, at much higher energy efficiency. In the 16-bit mode, it achieves over 600 TeraOps/s on an AMD MI300X GPU, while approaching 1 TeraOp/J. In the 1-bit mode, it breaks the 3 PetaOps/s barrier and achieves over 10 TeraOps/J on an NVIDIA A100 GPU. The beamforming library can be easily integrated into existing pipelines. We demonstrate its use for medical ultrasound and radio-astronomical instruments.
Paper Structure (16 sections, 9 equations, 7 figures, 3 tables)

This paper contains 16 sections, 9 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and related work
  3. The Tensor-Core Beamformer
  4. Tensor-core micro-benchmarks
  5. Complex number support
  6. Need for data reuse
  7. 1-bit arithmetic on tensor cores
  8. NVIDIA Hopper support
  9. Performance and energy efficiency
  10. Auto-tuning
  11. Roofline analysis
  12. Benchmarking
  13. Applications
  14. Computational ultrasound imaging
  15. Radio astronomy
  16. ...and 1 more sections

Figures (7)

  • Figure 1: 1-bit complex numbers and their binary representation. The representable values $-1\!-\!i$, $-1\!+\!i$, $1\!-\!i$, and $1\!+\!i$ are shown at the corners of the square, with binary values 00, 01, 10, and 11, respectively. The light gray circle represents the unit circle. Note that zero, i.e. $0\!+\!0i$, is not representable.
  • Figure 2: Auto-tuning results of ccglib matrix-matrix multiplication kernel. The measured performance and energy efficiency of each combination of tuning parameters is shown.
  • Figure 3: Roofline analysis of the ccglib matrix-matrix multiplication kernel. For each GPU, we show the roofline ceiling of the float16 and int1 (NVIDIA only) tensor cores, as well as the normal float32 cores for comparison.
  • Figure 4: Complex matrix-matrix multiplication benchmark results for \ref{['fig:benchmark_float16']} 16-bit data and \ref{['fig:benchmark_int1']} 1-bit data. The left panels show performance, while the right panels show energy efficiency.
  • Figure 5: Performance of beamforming for ultrasound. The number of voxels ranges from three orthogonal planes of $128\!\times\!128$ each, to the full $128^3$ data volume. The horizontal dash-dotted line indicates the minimum number of frames per second required for real-time performance.
  • ...and 2 more figures