This paper presents a programmable and scalable digital neuromorphic architecture based on 3D high-density memory integrated with logic tier for efficient neural computing. The proposed architecture consists of clusters of processing engines, connected by 2D mesh network as a processing tier, which is integrated in 3D with multiple tiers of DRAM. The PE clusters access multiple memory channels (vaults) in parallel. The operating principle, referred to as the memory centric computing, embeds specialized state-machines within the vault controllers of HMC to drive data into the PE clusters. The paper presents the basic architecture of the Neurocube and an analysis of the logic tier synthesized in 28nm and 15nm process technologies. The performance of the Neurocube is evaluated and illustrated through the mapping of a Convolutional Neural Network and estimating the subsequent power and performance for both training and inference.

Neurocube: a programmable digital neuromorphic architecture with high-density 3D memory

A single-chip microcomputer comprising: a first bus having a central processing unit and a cache memory connected therewith; a second bus having a dynamic memory access control circuit and an external bus interface connected therewith; a break controller for connecting the first bus and the second bus selectively; a third bus having a peripheral module connected therewith and having a lower-speed bus cycle than the bus cycles of the first and second buses; and a bus state controller for effecting a data transfer and a synchronization between the second bus and the third bus. The single-chip microcomputer has the three divided internal buses to reduce the load capacity upon the signal transmission paths so that the signal transmission can be accomplished at a high speed. Moreover, the peripheral module required to have no operation speed is isolated so that the power dissipation can be reduced.

Single-chip microcomputer

In one aspect of the present invention, a method for controlling the operation of a phase locked loop circuit is provided. The method is comprised of monitoring a frequency of a system clock, and a first signal is delivered in response to the detected frequency of the system clock being greater than a preselected setpoint. A second signal is delivered in response to the detected frequency of the system clock being less than a preselected setpoint. A first operating mode of the phase locked loop circuit is selected in response to receiving the first signal. The first mode of operation allows the phase locked loop circuit to synchronize with a clock signal in a first preselected range of frequencies. A second operating mode of the phase locked loop circuit is selected in response to receiving the second signal. The second mode of operation allows the phase locked loop circuit to synchronize with a clock signal in a second preselected range of frequencies.

Method and apparatus for reducing power consumption

A memory device with multiple clock domains. Separate clocks to different portions of the control circuitry create different clock domains. The different domains are sequentially turned on as needed to limit the power consumed. The turn on time of the domains is overlapped with the latency for the memory access to make the power control transparent to the user accessing the memory core. The memory device can dynamically switch between a fast and a slow clock depending upon the needed data bandwidth. The data bandwidth across the memory interface can be monitored by the memory controller, and when it drops below a certain threshold, a slower clock can be used. The clock speed can be dynamically increased as the bandwidth demand increases.

Power control system for synchronous memory device

A method for controlling the power used by a computer including the steps of measuring the operating characteristics of a central processor of the computer, determining when the operating characteristics of the central processor are significantly different than required by the operations being conducted, and changing the operating characteristics of the central processor to a level commensurate with the operations being conducted.

Adaptive power control

A method and apparatus for allowing a processor to invalidate an individual line of its internal cache while in a non-clocked low power state. The present invention includes circuitry for placing the processor in a reduced power consumption state. The present invention also includes circuitry for powering up the processor out of the reduced power consumption state to invalidate data in the cache in order to maintain cache coherency while in the reduced power consumption state.

Method and apparatus for invalidating a cache while in a low power state

A method and apparatus for reducing the power consumption of an integrated circuit when the core logic of the integrated circuit is in the quiescent or idle state. The method and apparatus includes a phase locked loop (PLL) circuit for generating an internal clock, wherein the frequency of the internal clock is at a predetermined multiple of the frequency of the global clock signal. When the integrated circuit is quiescent, the present invention provides circuitry which permits the internal clock to be slowed to a lower frequency or the internal clock to be frozen to reduce power consumption.

Method and apparatus for powering down an integrated circuit having a core that operates at a speed greater than the bus frequency

A 0.6–1.1 V, 84 Mb pipelined SRAM array design implemented in 14 nm FinFET CMOS technology is presented. Two array architectures featuring a high-density 0.0500 µm 2 6T SRAM bitcell and a 0.0588 µm 2 6T SRAM bitcell targeting low voltage operation are detailed. The high-density array design reaches 2.7 GHz at 1.1 V with 14.5 Mb/mm 2 bit density, while the low voltage optimized array can operate at 0.6 V, 1.5 GHz under typical process conditions. A capacitive charge-share transient voltage collapse write-assist circuit (CS-TVC) enables a 24% reduction in write energy compared to previous techniques by eliminating bias currents during operation. Technology and assist co-optimization enable $>$ 50 mV reduction in V $_{\rm MIN}$ and a 1.81× increase in density over a 22 nm design.

A 0.6 V, 1.5 GHz 84 Mb SRAM in 14 nm FinFET CMOS Technology With Capacitive Charge-Sharing Write Assist Circuitry

The growth of battery-powered mobile and wearable devices has increased the importance of low-power operation and cost in system-on-a-chip (SoC) design. Supply-voltage scaling is the predominant approach to active power reduction for SoC design, including voltage scaling for on-die memory given increasing levels of memory integration. SRAM can limit the minimum operating voltage (V MIN ) of a design, often leading to the introduction of separate voltage supplies for on-die memory. Additional supplies increase platform cost, and operating memory at higher voltage leads to increased power consumption. The introduction of trigate devices at the 22nm technology node delivered superior short channel effects and subthreshold slope relative to existing bulk planar device technology enabling reduction in threshold voltage within a fixed leakage constraint. Lower transistor V th , improvements to random device variability, and assist circuits to overcome device-size quantization enabled a >150mV reduction in SRAM V MIN [1]. At the 14nm technology node, FinFET device-size quantization remains a challenge for compact 6T SRAM bitcells with minimum-size transistors. Careful co-optimization between technology and design of memory-assist circuits is required in order to deliver dense, low-power memory operation at low voltages. In this paper, we present an 84Mb SRAM array design with wide-voltage-range operation in a 14nm logic technology featuring 2nd-generation FinFET transistors.

17.1 A 0.6V 1.5GHz 84Mb SRAM design in 14nm FinFET CMOS technology

A method and apparatus for powering down a microprocessor in a computer system. The method and apparatus includes a phase locked loop (PLL) circuit, wherein the phase locked loop generates bus clock signals for clocking the operations on the bus and core clock signals for clocking the core of the processor in response to global clock signal of the computer system. The microprocessor includes circuitry for processing data synchronous with the core clock signals. The method and circuit also includes circuitry for placing the processor in a reduced power consumption state in response to the execution of a power down instruction. In this manner, the computer system reduces power consumption.

James W. Conary

Papers

Method and apparatus for invalidating a cache while in a low power state

Method and apparatus for powering down an integrated circuit having a core that operates at a speed greater than the bus frequency

A 0.6 V, 1.5 GHz 84 Mb SRAM in 14 nm FinFET CMOS Technology With Capacitive Charge-Sharing Write Assist Circuitry

17.1 A 0.6V 1.5GHz 84Mb SRAM design in 14nm FinFET CMOS technology

Method and apparatus for powering down an integrated circuit transparently and its phase locked loop