Machine-learning accelerator (MLA) integrated circuit for extracting features from signals and performing inference computations

1. A machine-learning accelerator (MLA) integrated circuit for extracting features from signals and performing inference computations comprising:

a framework of finite state machine (FSM) kernels that are machine-learning algorithms implemented in hardware;

a kernel controller having mathematical structures implemented in hardware in communication with the framework of FSM kernels;

an arithmetic engine implemented in hardware in communication with the kernel controller to perform computations for the mathematical structures; and

a power management unit (PMU) that provides an idle-mode for the framework of FSM kernels, the kernel controller, and the arithmetic engine.

2. The MLA integrated circuit of claim 1 further including a peripheral interface bus for conveying data between a central processing unit (CPU) core and the framework of FSM kernels.

3. The MLA integrated circuit of claim 2 wherein the CPU core is implemented in hardware within the MLA integrated circuit.

4. The MLA integrated circuit of claim 3 wherein energy consumption per clock cycle for the CPU is around about 52 pJ for a power supply voltage of 0.7V.

5. The MLA integrated circuit of claim 1 wherein one of the FSM kernels is a Gaussian mixture model (GMM) FSM kernel.

6. The MLA integrated circuit of claim 5 wherein energy consumption of the GMM FSM kernel per clock cycle is around about 60 pJ for a power supply voltage of 0.7V.

7. The MLA integrated circuit of claim 1 wherein one of the FSM kernels is a support vector machine (SVM) FSM kernel.

8. The MLA integrated circuit of claim 7 wherein energy consumption of the SVM FSM kernel per clock cycle is around about 32 pJ for a power supply voltage of 0.7V.

9. The MLA integrated circuit of claim 1 wherein one of the FSM kernels is a discrete wavelet transform (DWT) FSM kernel.

10. The MLA integrated circuit of claim 9 wherein energy consumption of the DWT FSM kernel is around about 34 pJ for a power supply voltage of 0.7V.

11. The MLA integrated circuit of claim 1 wherein one of the FSM kernels is an active learning (AL) FSM kernel.

12. The MLA integrated circuit of claim 11 wherein energy consumption of the AL FSM kernel is around about 28 pJ for a power supply voltage of 0.7V.

13. The MLA integrated circuit of claim 1 wherein one of the FSM kernels is a general purpose (GP) FSM kernel that is user adaptable to generate user defined machine-learning algorithms.

14. The MLA integrated circuit of claim 1 wherein the mathematical structures include linear algebra constructs.

15. The MLA integrated circuit of claim 1 wherein the arithmetic engine includes a data computation unit (DCU) having a data path for multiplications, additions, and subtractions structured in stages with selectable operands, wherein the data path is selectably configurable by the kernel controller for each of the FSM kernels.

16. The MLA integrated circuit of claim 1 wherein the arithmetic engine includes a hardware-implemented coordinate rotation digital computer (CORDIC) for calculating exponential values, square root values, division values, and logarithmic values for data associated with the mathematical structures of the kernel controller.

17. The MLA integrated circuit of claim 1 wherein the arithmetic engine includes a hardware-implemented maximum and minimum block for determining maximum values and minimum values of data associated with the mathematical structures of the kernel controller.

18. The MLA integrated circuit of claim 1 further including a compression/decompression accelerator (CDA) implemented in hardware and coupled between a memory and the kernel controller for compressing data to be stored in the memory and for decompressing data retrieved from the memory.

19. The MLA integrated circuit of claim 18 wherein the memory is integral to the MLA integrated circuit.