Hardware architecture determination based on a neural network and a network compilation process
The described techniques provide for an automated process of hardware design using main applications (e.g., use cases) and a modeling heuristic of a compiler. Hardware may be optimized and designed based on the knowledge of the compiler for the hardware and firmware information. For instance, a user may define constraints (e.g., area constraints, power constraints, performance constraints, accuracy degradation constraints, etc.) and the changes of compilation heuristics and hardware parameters may be optimized to efficiently achieve the user defined constraints. Accordingly, hardware configuration parameters may be optimized based on the neural network's compilation process (e.g., actual compiler constraints) and optimization of power, performance, and area (PPA) constraints (e.g., user defined constraints). Specific neural processor (SNP) hardware may thus be designed based on the optimized hardware configuration parameters (e.g., via modeling heuristics of its compiler and main applications informed via user defined constraints or PPA constraints).
1. A method comprising:
generating hardware configuration parameters for a neural network hardware accelerator based on hardware constraints and compiler constraints, wherein the hardware configuration parameters include one or more memory sizes, one or more memory form factors, a number of multipliers, a topology of multipliers, or any combination thereof;
compiling a representation of a neural network architecture to produce a representation of an instruction set for implementing the neural network architecture based on the hardware configuration parameters and the compiler constraints;
simulating performance of the neural network architecture based on the hardware configuration parameters and the representation of the instruction set to produce a performance profile;
computing a performance metric based on the performance profile and power performance and area (PPA) constraints; and
iteratively updating the hardware configuration parameters for the neural network architecture based on the performance metric, simulating the performance of the neural network architecture, and recomputing the performance metric based on the updated hardware configuration parameters.
2. The method of claim 1 , wherein:
the performance profile includes two or more parameters from a set comprising algorithm performance, power consumption, memory bandwidth, and chip area.
3. The method of claim 1 , further comprising:
multiplying each parameter of the performance profile by a weight based on the PPA constraints; and
computing the performance metric based on a sum of each parameter of the performance profile multiplied by the corresponding weight, wherein the hardware configuration parameters are updated based on the sum.
4. The method of claim 1 , further comprising:
iterating between computing the performance metric and updating the hardware configuration to produce final hardware configuration parameters.
5. The method of claim 4 , wherein:
the iterating is based on simulated annealing.
6. The method of claim 4 , wherein:
the iterating is based on stochastic gradient descent.
7. The method of claim 1 , further comprising:
prompting a user to identify the PPA constraints; and
receiving the PPA constraints from the user based on the prompting.
8. The method of claim 1 , further comprising:
generating a register-transfer level (RTL) design for the neural network architecture based on the hardware configuration parameters.
9. The method of claim 1 , further comprising:
manufacturing a specific neural processor (SNP) based on the updated hardware configuration parameters.
10. The method of claim 1 , wherein:
the neural network architecture comprises a convolutional neural network (CNN).
11. The method of claim 1 , wherein:
the neural network architecture is configured to facial recognition tasks.
12. The method of claim 1 , further comprising:
selecting the hardware configuration parameters to be optimized based on the neural network architecture.
13. A method for hardware acceleration, comprising:
generating hardware configuration parameters for a neural network architecture based on hardware constraints;
compiling a representation of the neural network architecture to produce a representation of an instruction set for implementing the neural network architecture based on the hardware configuration parameters;
simulating performance of the neural network architecture based on the hardware configuration parameters and the representation of the instruction set to produce a performance profile;
computing a performance metric by weighting parameters of the performance profile based on power performance and area (PPA) constraints; and
iteratively updating the hardware configuration parameters for the neural network architecture based on the performance metric, simulating the performance of the neural network architecture, and recomputing the performance metric based on the updated hardware configuration parameters.
14. The method of claim 13 , wherein:
the hardware configuration parameters are generated based on compiler constraints, and the instruction set is compiled based on the compiler constraints.
15. An apparatus for hardware acceleration, comprising: a processor and a memory storing instructions and in electronic communication with the processor, the processor being configured to:
generate hardware configuration parameters for a neural network architecture based on hardware constraints and compiler constraints;
compute a performance metric based on power performance and area (PPA) constraints and a performance profile;
compile a representation of a neural network architecture to produce a representation of an instruction set for implementing the neural network architecture based on the hardware configuration parameters and the compiler constraints; and
simulate performance of the neural network architecture to produce a performance profile based on the hardware configuration parameters and the representation of the instruction set to produce a performance profile,
wherein the hardware configuration parameters for the neural network architecture are iteratively updated based on computing the performance metric, simulating the performance of the neural network architecture, and recomputing the performance metric based on the updated hardware configuration parameters.
16. The apparatus of claim 15 , further comprising:
a user interface configured to receive the PPA constraints from a user.
17. The apparatus of claim 15 , further comprising:
a design component configured to design a specific neural processor (SNP) based on the hardware configuration parameters.
18. The apparatus of claim 15 , wherein:
the decision making component takes compiler constraints corresponding to the policy making component as input.