Silicon Photonics Chiplet Accelerator for Low-Latency DNN Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Metallic-based interconnects in chiplet-based DNN accelerators face high latency, low bandwidth, and energy inefficiency, hindering performance and scalability due to challenges in data communication across chiplets.
Innovation Solution
A silicon photonics-based chiplet accelerator (SPACX) with a photonic network design enabling seamless single-chiplet and cross-chiplet multicast communications and a tailored dataflow that maximizes parallelism and energy efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If metallic-based interconnects are used for data communication in chiplet-based DNN accelerators, then the system can be implemented with conventional technology, but the communication latency increases and bandwidth decreases
Solution Approach 1:
The patent replaces metallic-based electrical interconnects with photonic interconnects using light to transmit data between chiplets. This substitution of the transmission medium fundamentally changes the communication mechanism, enabling simultaneous transmission of multiple data streams through wavelength division multiplexing, thereby reducing communication latency and increasing bandwidth while maintaining compatibility with chiplet-based architectures.
Solution Approach 2:
The patent changes the fundamental parameter of the transmission medium from electrical signals in metal interconnects to optical signals in photonic waveguides. This parameter change enables higher bandwidth and lower latency by utilizing the properties of light, such as high frequency and low attenuation, to overcome the limitations of metallic interconnects in scalable chiplet systems.
2Device complexity
If metallic-based interconnects are used for cross-chiplet communication, then the architecture can be simplified, but energy consumption increases due to long distance communication
Solution Approach 1:
The patent substitutes photonic transmission for metallic electrical interconnection to reduce energy consumption in long-distance cross-chiplet communication. Optical signals experience lower attenuation and require less energy for transmission over long distances compared to electrical signals in metal interconnects, thereby reducing overall system energy consumption while enabling scalable chiplet architectures.
Solution Approach 2:
The patent changes the transmission medium parameter from electrical to optical to address energy consumption issues. This parameter change exploits the low attenuation and high speed properties of light, enabling efficient long-distance communication between chiplets without the excessive energy consumption associated with metallic interconnects.
3Ease of operation
If photonic interconnects are used with equal bandwidth allocation, then communication between arbitrary nodes is balanced, but multicast and broadcast capabilities are not adequately supported
Solution Approach 1:
The patent implements dynamic bandwidth allocation in the photonic interconnect network, allowing the system to adapt bandwidth distribution based on communication patterns. This dynamic approach enables efficient support for multicast and broadcast operations by allocating bandwidth dynamically to groups of destinations rather than maintaining fixed equal allocation, thereby improving both versatility and operational efficiency.
Solution Approach 2:
The patent designs the photonic interconnect to perform multiple functions including unicast, multicast, and broadcast communication. By incorporating universal communication capabilities that can dynamically switch between different communication modes and patterns, the system achieves both balanced bandwidth allocation and robust support for diverse communication operations required by DNN inference workloads.
4Productivity
If dataflows are optimized for metallic-based interconnects, then data consumption in memory hierarchies closer to PEs is maximized, but communication distance becomes a primary concern and multicast is not promoted
Solution Approach 1:
The patent replaces metallic interconnect-based dataflows with photonic interconnect-based dataflows, fundamentally changing how data is transmitted across the system. This substitution eliminates the primary concern of communication distance by using optical transmission, which has lower attenuation and can maintain signal integrity over longer distances without requiring frequent data movement to nearby memory hierarchies, thereby enabling more flexible dataflow patterns.
Solution Approach 2:
The patent changes the transmission medium parameter from electrical to optical, which fundamentally alters the constraints on dataflow design. This parameter change removes the limitation of communication distance, allowing data to be efficiently distributed over longer distances through the photonic network, and enables multicast operations that were previously inefficient or impossible with metallic interconnects.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
SPACX achieves significant reductions in execution time and energy consumption by utilizing photonic interconnects for efficient data communication, supporting high parallelism and reducing energy consumption through multicast and broadcast capabilities.
Implementation Method 1
The photonic interconnects can provide adequate data communication support due to some superior properties like low latency, high bandwidth and energy efficiency
Implementation Method 2
A microring resonator (MRR) is provided in the interface, the MRR being configured to couple a fraction of the power of a specific wavelength belonging to the second wavelength set to the local waveguide
Data Source
AI summary
A computer architecture has a global optical waveguide, a buffer having memory space having a memory hierarchy above main memory for temporary data storage. A transmitter transmits a first optical signal with a first plurality of optical wavelengths on the global optical waveguide, and a second optical signal with a second plurality of optical wavelengths on the global optical waveguide. A receiver receives a second optical signal with a third plurality of optical wavelengths from the global optical waveguide. One or more local optical waveguide(s) are coupled to the global optical waveguide to receive all of the first plurality of optical wavelengths and a unique wavelength of the second plurality of optical wavelengths and transmit a unique wavelength of the third plurality of optical wavelengths. A plurality of chiplets are coupled to one of one or more local optical waveguides, each of the plurality of chiplets have a plurality of processing elements each receiving one of the first plurality of optical wavelengths and one of the second plurality of optical wavelengths from the local optical waveguide and transmitting one of the third plurality of optical wavelengths to local optical waveguide.