AQB8: Energy-Efficient Ray Tracing Accelerator through Multi-Level Quantization

Yen-Chieh Huang, Chen-Pin Yang, Tsung Tai Yeh
The 52nd Annual International Symposium on Computer Architecture (ISCA), June 21–25, 2025, Tokyo, Japan.

[Paper Link] [Poster Link] [PPT Link]

Introduction

AQB8 is a ray tracing accelerator designed to operate on quantized BVH trees using a novel multi-level quantization technique. This approach enables ray traversal to use low-bit integer arithmetic, significantly reducing reliance on FP32 operations. The hardware is simpler and more area- and energy-efficient.

Key benefits:

• 70% reduction in DRAM accesses
• 49% reduction in energy consumption
• 27% reduction in hardware area
• 1.82× performance speedup over modern GPU RT accelerators

Architectural modifications from a (a) baseline RT accelerator to our (b) quantized RT accelerator. The hardware block diagrams for the BOX and QBOX units are detailed in (c) and (d), respectively.

AQB8 on Vulkan-Sim

We provide a Docker image to run AQB8 on Vulkan-Sim.

Make sure Docker is installed before proceeding.

Quick Start

Run with the included script:

chmod +x docker_run.sh
./docker_run.sh

docker_run.sh example:

#!/bin/bash

# docker_run.sh - Run the AQB8 Vulkan-Sim Docker container

IMAGE="ycpin/aqb8-vulkan-sim"
CONTAINER_NAME="aqb8-vulkan-sim-container"

echo "Pulling Docker image: $IMAGE"
docker pull "$IMAGE"

echo "Running Docker container: $CONTAINER_NAME"
docker run -it --rm --name "$CONTAINER_NAME" "$IMAGE"

Inside the Container

The image includes six implementations of AQB8 on Vulkan-Sim, covering baseline, compressed, and quantized (AQB8) versions, each with 2-wide and 6-wide BVH trees.

Each folder corresponds to a Git branch containing a specific implementation:

/home/                                 <branch-name>
├── vulkan-sim-baseline-2wide/       ← baseline-2wide
├── vulkan-sim-baseline-6wide/       ← baseline-6wide
├── vulkan-sim-compress-2wide/       ← compress-2wide
├── vulkan-sim-compress-6wide/       ← compress-6wide
├── vulkan-sim-quantized-2wide/      ← AQB8-2wide
└── vulkan-sim-quantized-6wide/      ← AQB8-6wide

Branch Structure and Implementations

Each implementation is maintained in its own dedicated Git branch, named to match the corresponding folder. You can switch between them using:

git checkout <branch-name>

For example, to switch to the AQB8 2-wide simulation environment:

git checkout AQB8-2wide

Additionally, the hardware code for the AQB8 accelerator itself is available in a separate branch named RTL. To access the hardware design files, you can checkout this branch:

git checkout RTL

Usage Instructions

Navigate to one of the simulation folders, e.g.:

cd /home/vulkan-sim-quantized-2wide/vulkan-sim-root/

Step 1: Load Environment Variables

source embree-3.13.4.x86_64.linux/embree-vars.sh

Step 2: Build Vulkan-Sim + Mesa

cd vulkan-sim/
source setup_environment

cd ../mesa-vulkan-sim/
meson --prefix="${PWD}/lib" build -Dvulkan-drivers=swrast -Dgallium-drivers=swrast -Dplatforms=x11 -D b_lundef=false -D buildtype=debug
ninja -C build/ install  # May fail the first time — this is expected

export VK_ICD_FILENAMES=${PWD}/lib/share/vulkan/icd.d/lvp_icd.x86_64.json

cd ../vulkan-sim/
make -j

cd ../mesa-vulkan-sim/
ninja -C build/ install  # This second build should succeed

Step 3: Build the Ray Tracing App

cd ../RayTracingInVulkan/
./build_linux.sh

Step 4: Run Simulations

cd ..
bash rt_program_128.sh  # Resolution: 128x128
bash rt_program_256.sh  # Resolution: 256x256

Output logs and results are located in: /RayTracingInVulkan/build/linux/bin/

Citation

If you use AQB8 in your research, please cite our ISCA 2025 paper:

@inproceedings{huang2025aqb8,
  title     = {AQB8: Energy-Efficient Ray Tracing Accelerator through Multi-Level Quantization},
  author    = {Yen-Chieh Huang and Chen-Pin Yang and Tsung Tai Yeh},
  booktitle = {Proceedings of the 52nd Annual International Symposium on Computer Architecture (ISCA)},
  year      = {2025},
  address   = {Tokyo, Japan}
}

Contact

For questions or collaborations, please open an issue or contact the authors directly.