Multicycle paths : The architectural perspective


Definition of multicycle paths: By definition, a multi-cycle path is one in which data launched from one flop is allowed (through architecture definition) to take more than one clock cycle to reach to the destination flop. And it is architecturally ensured either by gating the data or clock from reaching the destination flops. There can be many such scenarios inside a System on Chip where we can apply multi-cycle paths as discussed later. In this post, we discuss architectural aspects of multicycle paths. For timing aspects like application, analysis etc, please refer Multicycle paths handling in STA.

Why multi-cycle paths are introduced in designs: A typical System on Chip consists of many components working in tandem. Each of these works on different frequencies depending upon performance and other requirements. Ideally, the designer would want the maximum throughput possible from each component in design with paying proper respect to power, timing and area constraints. The designer may think to introduce multi-cycle paths in the design in one of the following scenarios:
      
       1)      Very large data-path limiting the frequency of entire component: Let us take a hypothetical case in which one of the components is to be designed to work at 500 MHz; however, one of the data-paths is too large to work at this frequency. Let us say, minimum the data-path under consideration can take is 3 ns. Thus, if we assume all the paths as single cycle, the component cannot work at more than 333 MHz; however, if we ignore this path, the rest of the design can attain 500 MHz without much difficulty. Thus, we can sacrifice this path only so that the rest of the component will work at 500 MHz. In that case, we can make that particular path as a multi-cycle path so that it will work at 250 MHz sacrificing the performance for that one path only.
     
     2)      Paths starting from slow clock and ending at fast clock: For simplicity, let us suppose there is a data-path involving one start-point and one end point with the start-point receiving clock that is half in frequency to that of the end point. Now, the start-point can only send the data at half the rate than the end point can receive. Therefore, there is no gain in running the end-point at double the clock frequency. Also, since, the data is launched once only two cycles, we can modify the architecture such that the data is received after a gap of one cycle. In other words, instead of single cycle data-path, we can afford a two cycle data-path in such a case. This will actually save power as the data-path now has two cycles to traverse to the endpoint. So, less drive strength cells with less area and power can be used. Also, if the multi-cycle has been implemented through clock enable (discussed later), clock power will also be saved.

Implementation of multi-cycle paths in architecture: Let us discuss some of the ways of introducing multi-cycle paths in the design:

      1)      Through gating in data-path: Refer to figure 1 below, wherein ‘Enable’ signal gates the data-path towards the capturing flip-flop. Now, by controlling the waveform at enable signal, we can make the signal multi-cycle. As is shown in the waveform, if the enable signal toggles once every three cycles, the data at the end-point toggles after three cycles. Hence, the data launched at edge ‘1’ can arrive at capturing flop only at edge ‘4’. Thus, we can have a multi-cycle of 3 in this case getting a total of 3 cycles for data to traverse to capture flop. Thus, in this case, the setup check is of 3 cycles and hold check is 0 cycle.
Figure 1: Introducing multicycle paths in design by gating data path



    Now let us extend this discussion to the case wherein the launch clock is half in frequency to the capture clock. Let us say, Enable changes once every two cycles. Here, the intention is to make the data-path a multi-cycle of 2 relative to faster clock (capture clock here). As is evident from the figure below, it is important to have Enable signal take proper waveform as on the waveform on right hand side of figure 2. In this case, the setup check will be two cycles of capture clock and hold check will be 0 cycle.
   
   
When the launch clock is half in frequency, it is better to make the path a multicycle of 2 because data will anyways be launched once every few cycles.
Figure 2: Introducing multi-cycle path where launch clock is half in  frequency to capture clock


        2) Through gating in clock path: Similarly, we can make the capturing flop capture data once every few cycles by clipping the clock. In other words, send only those pulses of clock to the capturing flip-flop at which you want the data to be captured. This can be done similar to data-path masking as discussed in point 1 with the only difference being that the enable will be masking the clock signal going to the capturing flop. This kind of gating is more advantageous in terms of power saving. Since, the capturing flip-flop does not get clock signal, so we save some power too.
    
Figure 3: Introducing multi cycle paths through gating the clock path
      Figure 3 above shows how multicycle paths can be achieved with the help of clock gating. The enable signal, in this case, launches from negative edge-triggered register due to architectural reasons (read here). With the enable waveform as shown in figure 3, flop will get clock pulse once in every four cycles. Thus, we can have a multicycle path of 4 cycles from launch to capture. The setup check and hold check, in this case, is also shown in figure 3. The setup check will be a 4 cycle check, whereas hold check will be a zero cycle check.

Pipelining v/s introducing multi-cycle paths: Making a long data-path to get to destination in two cycles can alternatively be implemented through pipelining the logic. This is much simpler approach in most of the cases than making the path multi-cycle. Pipelining means splitting the data-path into two halves and putting a flop between them, essentially making the data-path two cycles. This approach also eases the timing at the cost of performance of the data-path. However, looking at the whole component level, we can afford to run the whole component at higher frequency. But in some situations, it is not economical to insert pipelined flops as there may not be suitable points available. In such a scenario, we have to go with the approach of making the path multi-cycle.

References:



6 comments:

  1. Hi,

    I have doubt in diagram 3. which flipflop is used while generating enable signal. Is it D or T? if it is D with always high input how enable signal goes down in once it gets set on negedge. If it is T flipflop why it didnt toggle in between two pulses of present enable signal when negedge of capture flop arrives.
    If it is D with

    ReplyDelete
    Replies
    1. Hi Nikhil

      The enable generation logic does not need to be so simple. It may be state machine working on positive edge and then a negative edge-triggered D flip-flop acting as a pipeline as well. :-)

      And you are right, the waveforms should reflect the design. I haven't shown the complete design as the intent was to show how a multicycle path is formed.

      Delete
  2. Why the setup check is 3 cycles and hold check is only 0 cycles?

    ReplyDelete
    Replies
    1. This is because the timing path between launch and capture flops is enabled every third cycle. So, if setup check is 3 cycles, hold check is 0 cycle. In other words, it is governed by how you write the RTL code; i.e. it is governed by architecture.

      Delete
  3. In your last paragraph you mention that pipelining "eases the timing at the cost of performance of the data-path". How is performance negatively impacted when that path was going to take two cycles anyway. You shouldn't be negatively impacting the performance.

    One reason that I think you should have highlighted is why exactly some paths cannot be pipelined and are forced to be multicycled instead. Pipelining adds the setup and clk-to-q delays of the newly added pipeline flop to your existing path. So if the combo logic needed for that path cannot handle the added setup and clk-to-q delays of pipeline flop then you have to multicycle.

    ReplyDelete
    Replies
    1. Hi

      With multicycling, you have to reduce the toggling of the data. This is not the case with pipelining. The path will take 2 cycles, but still it can toggle at same rate.

      Delete