This application is related to German Application Number 07100538.3, filed Jan. 15, 2007 and related to Ser. No. 11/850,745 filed contemporaneously with this Application.
The present invention relates to clock gating in a synchronous circuit including multiple register stages.
In a synchronous circuit within a microprocessor, a clock signal is used to define a time reference for the movement of data within that circuit. Such a synchronous circuit includes multiple register stages in what is commonly referred to as a pipeline.
Modern microprocessors suffer a huge power drawback, where power is a combination of static and dynamic contributions, where the static contribution is approximately proportional to the silicon area of a macro. One approach to reduce dynamic power is to switch off the clock signal of not required register stages within the synchronous circuit. This procedure is also known as clock gating. By clock gating, the power dissipated in the clock mesh is reduced as well as data switching power is reduced. The latter is due to the fact that the outputs of the registers are constant if there is no clock signal.
To maximize the benefit of clock gating clock activation or deactivation is performed in a small grain—in the extreme case on a single cycle basis.
A simple clock gating scheme is known from Li et Al, “Deterministic Clock Gating for Microprocessor Power Reduction”; Proceedings of the 9th International Symposium on High Performance Computer-Architecture, 2003, pp. 113-122, wherein a register to be clock gated is connected with the output of an AND-gate whose first input is fed with the clock signal and whose second input is fed with a clock activation signal. A problem in applying this approach for a pipeline is that for each stage a clock activate signal plus a clock signal is needed that are synchronously fed to AND-gates dedicated to the stages.
A common solution for this problem is to generate local clock signals by local clock buffers (LCB) within the stages and to provide a clock activate signal for the first stage of a pipeline and to propagate this clock activate signal parallel to the data signal through the pipeline. The clock activation signal when propagated from stage to stage activates the LCB of a stage it reaches which LCB activates the particular stage by clocking a data register in order to forward a data signal stored within that data register to the next stage. To propagate the clock activation signal from stage to stage synchronous with the data signal, control registers arranged in parallel to the data registers of the pipeline are used. Since the LCB of a stage that activates the data register belonging to that stage would continue to activate the clock whenever the corresponding clock activation signal stored in the control register of the previous stage is high, the control registers of the stages have to be clocked at least twice. Thereby the first clocking is to latch a clock activation signal into the control register in order to activate the LCB of the data register and the second clocking is to reset the clock activation signal in that control register in order to stop activation of the LCB. Due to this, the clock activation signal cannot be propagated through the pipeline within the same clock domain as the data signal as long as the wider and therefore higher power data registers are clocked only once. According to the state of the art, a second LCB per stage is required that forms the clock domain for the clock activation signal propagation within the pipeline.
State-of-the-art circuitries as the circuitry 1 shown in
A main drawback of the synchronous circuit according to the state of the art is that in each stage of the pipeline, two LCBs are needed, one for the clock domain of the data registers for the data signal and a second one for the clock domain of the control registers. This is disadvantageous because LCBs are highly complex and large circuitries with high silicon area requirements and high active as well as leakage power consumption.
It is therefore an object of the invention to provide a method to reduce power consumption within a clock gated synchronous circuit, plus to provide a clock gated synchronous circuit to be used to perform such a method.
The first object of the invention is met by a method to reduce power consumption within a clock gated synchronous circuit, wherein said synchronous circuit comprises at least two successive stages, wherein each stage if activated propagates a data signal cycle by cycle to a succeeding stage. Said method according to the invention comprises the steps of:
deriving a local clock activation signal from an external clock activation signal which external clock activation signal if high specifies that a corresponding data signal is valid and has to be propagated through the synchronous circuit by activating a particular stage and if low specifies that a corresponding data signal is invalid and clock gating has to be performed by deactivating a particular stage, wherein said local clock activation signal changes its value every cycle the external clock activation signal is high,
propagating the data signal and the local clock activation signal synchronously cycle by cycle from a particular stage to a succeeding stage by activating said particular stage whenever a local clock activation signal at the particular stage by derivation from the clock activation signal or by propagation through the synchronous circuit changes its value between two successive cycles, in order to propagate the data signal and the local clock activating signal within the same clock domain through the clock gated synchronous circuit.
Thereby it is important to mention that it is free to define, whether the external clock activation signal is high or low at high or low voltage or vice versa.
The foregoing, together with other objects, features, and advantages of this invention can be better appreciated with reference to the following specification, claims and drawings, with:
The idea behind the method according to the invention is to propagate a signal activating the particular stages of a pipeline within the same clock domain as the data signal itself in order to save the LCBs for the clock activation signal clock domain within the pipeline. To do so, the clock activation signal used in common microprocessors cannot be used. A new local clock-activation signal derived from the common external clock activation signal is needed showing attributes allowing propagating it within the same clock domain as the data signal. The solution is a local clock activation signal that by changing its value from one cycle to another indicates that a particular stage has to be activated. In this way, the local clock activation signal only needs to be clocked once per cycle, i.e. within the same clock domain as the data signal. Hence, the local clock activate signal can be latched in the same clock domain as the data signal. The benefit of this approach is two-fold: First, any data and control register is clocked only a single time per data signal and pipeline stage. Second, a single clock domain can be used for the local clock activation and the data signals, i.e. the LCB commonly used for the clock domain of the clock activation signal can be saved, reducing silicon area, cost and power consumption.
It is thus an additional advantage of the method according to the invention, that particularly within pipeline architectures with short pipelines, i.e. with few stages, LCBs can be saved compared to the state of the art. Also within large pipelines, power savings are significant.
According to a preferred embodiment of the method according to the invention, the derivation of the local clock activation signal from the external clock activation signal is performed in a first stage of the synchronous circuit, wherein a propagation of the local clock activation signal and of the data signal at the first stage is controlled by the external clock activation signal.
A particularly preferred embodiment of the method according to the invention comprises the steps of:
According to another preferred embodiment of the method according to the invention, generating a local clock activation signal within said first stage that changes its value whenever said first stage is activated is performed by a first stage control register activated synchronously to a first stage data register by a first stage LCB activated by the external clock activation signal, wherein the output of the first stage control register is used as an input for a second stage control register, and wherein the output of the first stage control register is inverted and re-used as an input of the first stage control register too.
According to an additional preferred embodiment of the method according to the invention, activating a following stage when the local clock activation signal propagated to said stage changes its value compared to the last local clock activation signal propagated to said stage within the last cycle a propagation occurred is performed by using one control register per stage to latch the local clock activation signal cycle by cycle when the particular stage is activated and wherein a changing of the value of the local clock activation signal is determined by comparing the input of the control register of said particular stage with the output of said control register of said stage.
According to a particularly preferred embodiment of the method according to the invention, comparing the input of the control register of a particular stage with the output of said control register of said stage is performed by an XOR-gate with the input and the output of the control register of the particular stage as inputs of said XOR-gate respectively, and the output of said XOR-gate being used to activate the LCB of that particular stage clocking the control register and the data register synchronously in order to propagate the local clock activation signal and the data signal synchronously to the next stage.
A main idea of this embodiment of the method according to the invention is to activate a stage comprising a LCB, a control and a data register by arranging a XOR-gate between the input and the output of the data register wherein the output of the XOR-gate is used to activate the LCB of said stage. An advantage of this embodiment is that it is cheap and easy to realize.
A second aspect of the invention concerns a clock gated synchronous circuit to perform a method mentioned above. The clock gated synchronous circuit comprises at least two successive stages of at least a control register, a data register and a LCB each, wherein each stage if activated propagates a data signal stored within the data register cycle by cycle to a data register of a succeeding stage. Said clock gated synchronous circuit according to the invention comprises means to derive a local clock activation signal from an external clock activation signal which external clock activation signal if high specifies that a corresponding data signal is valid and has to be propagated through the synchronous circuit by activating a particular stage and if low specifies that a corresponding data signal is invalid and clock gating can be performed by deactivating a particular stage, wherein said local clock activation signal changes its value every cycle the external clock activation signal is high, means to propagate the data signal and the local clock activation signal synchronously cycle by cycle from a particular stage to a succeeding stage by activating said particular stage whenever a local clock activation signal at the particular stage by derivation from the clock activation signal or by propagation through the synchronous circuit changes its value between two successive cycles, in order to propagate the data signal and the local clock activating signal within the same clock domain through the clock gated synchronous circuit.
According to a preferred embodiment of the clock gated synchronous circuit according to the invention, the means to derive a local clock activation signal from an external clock activation signal are arranged within a first stage of said clock gated synchronous circuit, wherein said first stage also comprises means to activate said first stage by said external clock activation signal, in order to control a propagation of the local clock activation signal and of the data signal at the first stage by the external clock activation signal.
According to an additional preferred embodiment of the clock gated synchronous circuit according to the invention, the means to activate the first stage by the external clock activation signal comprise a LCB activated by the external clock activation signal, wherein the LCB when activated clocks the control register and the data register of the first stage.
According to a particularly preferred embodiment of the clock gated synchronous circuit according to the invention, the means to derive a local clock activation signal from an external clock activation signal comprise a first stage control register activated synchronously to a first stage data register by a first stage LCB, wherein the output of the first stage control register is connected with the input of a second stage control register, and wherein the output of the first stage control register via an inverter is connected with the input of the first stage control register too.
Said means to derive a local clock activation signal from an external clock activation signal generate a local clock activation signal within said first stage that changes its value whenever said first stage is activated. Doing so ensures that each time the external clock activation signal activates the first stage LCB the local clock activation signal propagated from the first stage control register to the second stage control register changes its value.
A preferred embodiment of the clock gated synchronous circuit according to the invention comprises means to activate a following stage when the local clock activation signal propagated to said following stage changes its value compared to the last local clock activation signal propagated to said stage within the last cycle a propagation occurred.
According to a preferred embodiment of the clock gated synchronous circuit according to the invention, the means to activate a following stage when the local clock activation signal propagated to said following stage changes its value compared to the last local clock activation signal propagated to said stage within the last cycle a propagation occurred comprise a control register per stage to latch the local clock activation signal cycle by cycle when the particular stage is activated and means to compare the input of the control register of the particular stage with the output of said control register of said stage.
According to an additional preferred embodiment of the clock gated synchronous circuit according to the invention, the means to compare the input of the control register of the particular stage with the output of said control register of said stage comprise an XOR-gate connected with the input and the output of the control register of the particular stage as inputs of said XOR-gate respectively, wherein the output of said XOR-gate is connected with the LCB of that particular stage to activate the control register and the data register of the particular stage synchronously when activating the LCB by the output of the XOR-gate, in order to propagate the local clock activation signal and the data signal synchronously to the next stage.
According to a particularly preferred embodiment of the clock gated synchronous circuit according to the invention, the first stage of the circuit comprises a LCB activated by the external clock activation signal, a control register and a data register, both connected with the LCB of the first stage in order to be activated when the LCB is activated by the external clock activation signal, wherein the first stage control register with its output is connected with the input of a control register of a succeeding stage and the output of the first stage data register is connected with the input of the data register of the succeeding stage and wherein the output of the first stage control register is also connected with an inverter that is connected with the input of the first stage control register, in order to generate the local clock activation signal that whenever the first stage is activated by the external clock activation signal being high changes its value and is propagated to the next stage synchronously to a data signal, and wherein any following stage comprises a control register connected with a control register of a previous and/or a succeeding stage, a data register connected with a data register of a previous and/or a succeeding stage and a LCB connected with its output with the control and data register of the particular stage in order to activate them synchronously when activated, wherein the LCB further is connected with an output of an XOR-gate whose inputs are connected with the input and the output of the control register of the particular stage respectively in order to activate the LCB whenever the local clock activation signal propagated to said stage changes its value compared to a previous cycle an activation occurred.
The idea is to propagate a local clock activation signal instead of the common clock activation signal through the pipeline. In this way, the differential clock activation signal only needs to be clocked once. Hence, the local clock activate signal can be latched in the same clock domain as the data signal. The benefit of this approach is two-fold: First, any register and clock signal is clocked only a single time per data word and pipeline stage. Second, a single clock domain can be used for clock activation and data signals.
An example of a clock gated synchronous circuit 10 applying this principle is shown in
If the local clock activation signal dact_0 of the first stage I in
While the present invention has been described in detail, in conjunction with specific preferred embodiments, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art in light of the foregoing description. It is therefore contemplated that the appended claims will embrace any such alternatives, modifications and variations as falling within the true scope and spirit of the present invention.
Number | Date | Country | Kind |
---|---|---|---|
20 2007 100 538 U | Jan 2007 | DE | national |
Number | Name | Date | Kind |
---|---|---|---|
20030131270 | Abernathy et al. | Jul 2003 | A1 |
Number | Date | Country | |
---|---|---|---|
20080169841 A1 | Jul 2008 | US |