Paper title:
NEUTRAMS: Neural Network Transformation and Co-design under Neuromorphic Hardware Constraints
Publication:
MICRO’16
Problem to solve:
Another computing paradigm, neuromorphic computing, is emerging as a hot scheme with associated design and fabrication of neuromorphic chips. To address the difficult of programming for neuromorphic computing, this paper proposed a systematic methodology with a set of tools. The tool is necessary to be designed to bridge the gap between applications and neuromorphic hardware.
Major contribution:
-
This paper proposed a hardware-independent representation layer of neural network to describe high-level NN models to satisfy the hardware constraints of a neuromorphic chip.
-
This paper designed a training-based transformation algorithm to transform an existing NN described by the above representation into its counterpart under the target hardware’s constraints.
-
This paper implemented a configurable, clock-driven simulator to achieve performance and power simulation on the microarchitecture level, supporting different types of neuromorphic cores and memory technologies.
-
This paper proposed an optimized strategy to map the trained NNs onto neuromorphic hardware, using graph partition algorithm to put densely communicating neurons together.
-
This paper implemented such a toolchain for a real CMOS neuromorphic processor and for a processing-in-memory architecture design for ANNs, guiding the optimization of hardware design.
Lessons learnt:
-
Learning hint from this paper, employing a neuromorphic technique to form the general system for diverse NNs, such as SNNs, CNNs, RNNs and MLPs, may come a high-level value. As such, designing a novel architecture for a large chip domain, break through the fundamental architecture such von Neumann architecture may bring an insight into the designing space.
-
From an abstract idea to a visual general implements, a set of toolchains or evaluators is essential process. Of these, toolchains can help achieve the mapping among preliminary idea and hardware/software. Subsequently, the error-rates and hardware constraints and consumptions may be the target, worthy of achieving.