130 lines
5.2 KiB
Markdown
130 lines
5.2 KiB
Markdown
# SPACX hybrid Network on Chip (NoC) Simulation
|
|
|
|
This project simulates the Optical plus Electrical Network on Chip (NoC) architecture SPACX using SystemC TLM 2.0
|
|
|
|
## Prerequisites
|
|
|
|
- SystemC (version 2.3.3 or later)
|
|
- C++ compiler supporting C++17 or later (e.g., GCC 7+ or Clang 5+)
|
|
- Make
|
|
|
|
## Setup
|
|
|
|
1. Install SystemC on your system.
|
|
2. Set the `SYSTEMC_HOME` environment variable to point to your SystemC installation directory:
|
|
|
|
```bash
|
|
export SYSTEMC_HOME=/path/to/your/systemc/installation
|
|
|
|
## Compilation
|
|
make
|
|
|
|
## Output
|
|
./out/simulation
|
|
|
|
or
|
|
|
|
make run
|
|
|
|
|
|
## Project directory structure
|
|
|
|
SPACX
|
|
- src
|
|
- GB
|
|
- Global_Buffer.h
|
|
- Global_Buffer.cpp
|
|
- Interposer
|
|
- Interposer_Interface.h
|
|
- Interposer_Interface.cpp
|
|
- Chiplet
|
|
- Chiplet_Interface.h
|
|
- Chiplet_Interface.cpp
|
|
- Optical_Electrical
|
|
- O_E_converter.h
|
|
- O_E_converter.cpp
|
|
- PE
|
|
- Processing_Element.h
|
|
- Processing_Element.cpp
|
|
- Utils
|
|
- configuration.h
|
|
- noc_logger.h
|
|
- noc_logger.cpp
|
|
- token_manager.h
|
|
- scripts
|
|
- input_feature.py
|
|
- weight_kernal.py
|
|
- out
|
|
- obj
|
|
- main.cpp
|
|
- Makefile
|
|
- README.md
|
|
|
|
# Log files
|
|
|
|
The log files are in
|
|
- out/report.log : general communication like token transfer
|
|
- out/gb_comm.log : communication to and from global buffer
|
|
- out/interposer_comm.log : communication to and from interposer interface
|
|
- out/chiplet_comm.log : communication to and from chiplet interface
|
|
- out/oec_comm.log : communication to and from OEC
|
|
- out/pe_comm.log : communication to and from PE
|
|
|
|
# SPACX initial implementation
|
|
|
|
- The Global Buffer is initializing all the transactions
|
|
- In the beggining GB generates weights and input features as matrices and convert them to payloads
|
|
- The GB is connected to interposer interface to accept the payload through TLM blocking communication
|
|
- The address is the wavelegth, which is encoded into the payload and transmitted
|
|
- The interposer interface simply diverts the payload to chiplet interface again as TLM blocking communication
|
|
- The chiplet interface is connected to all PEs in a chiplet through the optical to electrical converter
|
|
- The chiplet interface copies the data to the curresponding OEC based on the address that is the wavelength
|
|
- Once the data is in the OEC, they are simply forwarding them to the PEs using the TLM blocking communication
|
|
- In PEs the data is stored into 2 fifos, the weight buffer and the buffer for input features
|
|
- The PE then multiplies it and gets the Psum stored into the accumulation buffer
|
|
- once the psum is ready, PE weights for the token, and then gives the data to the OEC
|
|
- OEC then transfers the data to the chiplet interface, and then to interposer interface in backward
|
|
- This way the Psums are back at global buffer through multiple TLM blocking transport
|
|
|
|
# SPACX architecture and its operation
|
|
|
|
1. SPACX Overview:
|
|
SPACX leverages photonic interconnects for efficient data communication in chiplet-based architectures, focusing on the needs of DNN inference, such as broadcast communication. It uses a hierarchical design:
|
|
- Global waveguide: Connects all chiplets.
|
|
- Local waveguide: Exists within each chiplet for intra-chiplet communication.
|
|
2. Key Components:
|
|
Chiplets and Processing Elements (PEs):
|
|
- Each chiplet has multiple PEs (e.g., 8 per chiplet in the example).
|
|
- PEs perform computations like Multiply-and-Accumulate (MAC) for DNN workloads.
|
|
- Each PE has:
|
|
- Two receivers for incoming broadcast data (local and global).
|
|
- One transmitter for sending computed results back to the global buffer (GB).
|
|
Wavelengths and Waveguides:
|
|
- Data transmission uses wavelength-division multiplexing (WDM).
|
|
- Wavelength allocation:
|
|
- Global (cross-chiplet) wavelengths: Used for broadcasting the same data to corresponding PEs across all chiplets.
|
|
- Local (single-chiplet) wavelengths: Used for intra-chiplet broadcasts and communication between chiplet PEs and the GB.
|
|
Optical Splitters and Filters:
|
|
- Optical splitters: Split light (data) across multiple destinations by controlling the split ratio. Used for broadcasts.
|
|
- Optical filters: Select or forward specific wavelengths for targeted communication.
|
|
3. Communication Mechanisms:
|
|
A. Cross-Chiplet Broadcast (Global Waveguide):
|
|
- Example: Data for PE0 across all chiplets is sent using wavelength λ0.
|
|
- Process:
|
|
- The GB modulates data onto λ0.
|
|
- Splitters in the global waveguide distribute power to each chiplet's local waveguide.
|
|
- Local waveguides deliver the data to PE0 on each chiplet.
|
|
B. Single-Chiplet Broadcast (Local Waveguide):
|
|
- Example: Data for all PEs within a single chiplet (e.g., Chiplet0) is sent using wavelength λ8.
|
|
- Process:
|
|
- The GB modulates data onto λ8.
|
|
- Optical filters direct all power of λ8 from the global waveguide to the local waveguide.
|
|
- Splitters on the local waveguide evenly distribute data to all PEs within Chiplet0.
|
|
C. PE-to-GB Communication:
|
|
- Each chiplet's PEs share a single wavelength for sending computed results back to the GB (e.g. λ8 for Chiplet0).
|
|
- Token-based arbitration ensures orderly transmission:
|
|
- A token is passed sequentially among PEs, granting transmission access to one PE at a time.
|
|
|
|
|
|
|
|
|