HWComponents-NeuroSim provides energy, area, and leakage models for NeuroSim[1] components. It enables estimation for analog Compute-In-Memory (CiM) Array components and various peripheral components.
We also allow conversions for NVMExplorer[2]/NVSim[3] style cell files to Neurosim processing-in-memory / compute-in-memory arrays.
We encourage you read section 6 of the NeuroSim Manual in the DNN_NeuroSim_V1.3/Documents folder.
We model CiM array energy/area with four pieces:
- Row Activation: This is the path for inputs to activate a array. An input voltage will actiavte a wordline, or row of memory cells. Multiple input voltages or longer-duration inputs can feed higher resolution inputs.
- Column Activation: This is the path for outputs to be read out of a array. Once input voltages have activated a wordline, they will feed current onto bitlines, or columns. Column activation is the energy required to get this current out and to precharge the wordline before reading.
- Analog-Digital-Conversion: This is the circuit that reads an analog value from the array and converts to a digital signal for output processing.
- Cell: This a memory cells in the array.
In short: An input activates a row, an output activates a column, a multiplication activates a cell, and an output read takes an ADC conversion.
Clone the repository and install with pip:
git clone --recurse-submodules https://github.com/Accelergy-Project/hwcomponents-neurosim.git
cd hwcomponents-neurosim
pip install .
# Check that the installation is successful
hwc --list | grep array_col_driversCell files must follow the NVMExplorer[2]/NVSim[3] format. See the cells/ directory for references. CiM arrays can be created with any user-defined cell.
We support four components for estimating CiM array energy.
- array_row_drivers: These activate rows.
- array_col_drivers: These activate columns.
- array_adc: These read out analog values into digital signals.
- memory_cell: This is a cell in a array.
CiM components take the following arguments.
- tech_node: The technology node in m
- row: The number of rows in the array
- cols: The number of columns in the array
- cols_active_at_once: The number of columns to be activated at a single time
- cell_config: The path to a NVMExplorer[2]/NVSim[3] cell file, or a sample cell to use.
- average_input_value: A value between 0 and 1 reflecting the average input being sent on rows. For example, if inputs are encoded as values from 0-4 with an average of 3, average_input_value is 0.75.
- average_cell_value: Like average_input_value, but with encoded weights. For example, if weights are encoded as values from 1-10 with an average of 7, average_cell_value is 0.7.
- sequential: A binary value. If true, rows are addressed and accessed one at a time. Otherwise, rows are to be activated in large blocks and not addressed. Setting this to TRUE can simulate a CiM memory.
- adc_resolution: This is the number of bits of the ADC used for readout. ADC is a flash ADC. To exclude the ADC and use your own, set adc_resolution to 0.
- read_pulse_width: Number of ns each read pulse lasts.
- voltage_dac_bits: Number of bits resolution for a voltage-based DAC on each row. Voltage DACs use a row switch connected to a power rail for each possible input value.
- temporal_dac_bits: Number of bits resolution for a temporal DAC on each row. Temporal DACs enode inputs as an amount of time the row stays high.
We support other digital components from NeuroSim. These components are useful in many places-- not just CiM!
Yep, it adds integers.
Class:
- intadder
Required parameters:
- tech_node: The technology node in m
- n_bits: Number of bits to add
Actions:
- add
An adder + a shift register to accumulate variable-precision values.
Class:
- shift_add
Required parameters:
- tech_node: The technology node in m
- n_bits: Number of bits to add
- shift_register_n_bits: Number of bits in the shift register
Actions:
- shift_add
A tree of adders to accmulate many values. Each level of the tree adds with higher n_bits to ensure no overflow.
Class:
- intadder_tree
Required parameters:
- tech_node: The technology node in m
- n_bits: Number of bits of a leaf adder. This is the minimum n_bits in the tree. Each additional level will add a bit.
- n_adder_tree_inputs: Number of values to add
Actions:
- add
A max-pooling unit that finds and outputs the maximum a set of values.
Class:
- max_pool
Required parameters:
- tech_node: The technology node in m
- n_bits: Number of bits for each input value
- pool_window: Number of values to compare
Actions:
- max_pool
An n-bit multiplexer
Class:
- mux
Required parameters:
- tech_node: The technology node in m
- n_bits: Number of bits for each input value
- n_inputs: Number of inputs to the mux
Actions:
- max_pool
A digital flip flop
Class:
- flip_flop
Required parameters:
- tech_node: The technology node in m
- n_bits: Number of flip flop bits
Actions:
- max_pool
Logic gates
Class:
- not_gate
- nand_gate
- nor_gate
Required parameters:
- tech_node: The technology node in m
Actions:
- read
[1]X. Peng, S. Huang, Y. Luo, X. Sun, and S. Yu, “DNN+NeuroSim: An End-to-End Benchmarking Framework for Compute-in-Memory Accelerators with Versatile Device Technologies,” in 2019 IEEE International Electron Devices Meeting (IEDM), Dec. 2019, p. 32.5.1-32.5.4. doi: 10.1109/IEDM19573.2019.8993491.
[2]L. Pentecost, A. Hankin, M. Donato, M. Hempstead, G.-Y. Wei, and D. Brooks, NVMExplorer: A Framework for Cross-Stack Comparisons of Embedded Non-Volatile Memories. 2021.
[3]X. Dong, C. Xu, Y. Xie, and N. P. Jouppi, “NVSim: A Circuit-Level Performance, Energy, and Area Model for Emerging Nonvolatile Memory,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 31, no. 7, pp. 994–1007, Jul. 2012, doi: 10.1109/TCAD.2012.2185930.