NEureka is a Deep Neural Network accelerator which exploits the Hardware Processing Engine (HWPE) paradigm [1] (https://hwpe-doc.rtfd.io) and is designed to be integrated in an open-source PULP cluster configuration in combination with the Heterogeneous Cluster Interconnect (HCI). It makes use of the open-source IPs 'hci', 'hwpe-ctrl', and 'hwpe-stream'.
In general NEureka has built-in HW supports the following features:
- Filters: 1x1, 3x3, depthwise
- Batch normalization
- ReLU
- Activation input bits: 8
- Weight bits: 2,3,4,5,6,7,8
- Activation output bits: 8,32
- Nr of input channels: arbitrary
- Nr of output channels: arbitrary
NEureka is a direct derivative of the NE16 design https://github.com/pulp-platform/ne16 .
The simulation infrastructure of NEureka uses QuestaSim.
To build the environment, you can run the following with QuestaSim in your PATH:
# fetch Bender (if not available), update the dependencies, and generate the scripts
make update-ips
# build the simulation environment
make hw-all
This version of NEureka relies on the https://github.com/pulp-platform/pulp-nnx library to generate simulation stimuli. You can fetch it in the model/deps folder as a submodule:
git submodule update --init
The pulp-nnx library has several Python requirements, such as PyTorch. Refer to model/requirements.txt for a list; you can install a Python VirtualEnv by running the create_venv.sh script and then entering the created environment:
source create_venv.sh
source venv/bin/activate
You also need a RISC-V GCC toolchain, i.e., riscv32-unknown-elf-gcc must be in your PATH.
You can generate stimuli with
make stimuli
To build the software generated test,
make sw-all
Finally, to run it:
# without QuestaSim GUI
make run
# with GUI
make run gui=1
There are several controllable parameters, such as filter size FS (1 or 3), input spatial size H_IN and W_IN, input channels K_IN, output channels K_OUT, depthwise conv (with FS=3) DW.
See the Makefile for a full list. All simulation commands should be augmented with modified parameters, e.g.,
make stimuli H_IN=7 W_IN=3 K_OUT=32 K_IN=32
make sw-all run H_IN=7 W_IN=3 K_OUT=32 K_IN=32 gui=0
- Arpan Suravi Prasad, ETH Zurich ([email protected])
- Francesco Conti, University of Bologna ([email protected])
See https://pulp-platform.github.io/neureka/dev/bench/
This repository makes use of two licenses:
- for all software: Apache License Version 2.0
- for all hardware: Solderpad Hardware License Version 0.51
For further information have a look at the license files: LICENSE.hw, LICENSE.sw
- A. S. Prasad, L. Benini and F. Conti, "Specialization meets Flexibility: a Heterogeneous Architecture for High-Efficiency, High-flexibility AR/VR Processing," 2023 60th ACM/IEEE Design Automation Conference (DAC), San Francisco, CA, USA, 2023, pp. 1-6, doi: 10.1109/DAC56929.2023.10247945.
- A. S. Prasad, M. Scherer, F. Conti, D. Rossi, A. Di Mauro, M. Eggimann, J. T. Gómez, Z. Li, S. Shakib Sarwar, Z. Wang, B. De Salvo, and L. Benini, "Siracusa: A 16 nm Heterogenous RISC-V SoC for Extended Reality with At-MRAM Neural Engine," IEEE Journal of Solid-State Circuits, 2024 (accepted), arXiv: https://arxiv.org/abs/2312.14750.