SSA in buffers

As mentioned before, buffers can be very power-hungry too. To get the most benefit out of signal gating, a signal with SSA should be gated before it propagates into buffers. When buffers are inserted along a wire to reduce interconnect delay, gating at the most upstream location reduces SSA in both the wire and the buffers. When the wire is driven by an output buffer, gating should be placed before the buffer. For a fully dedicated output network, instead of using one large output buffer to drive the whole output network, it is better to use a smaller buffer for each dedicated interconnect and its receiving DPU while gating the signal with SSA before each buffer. Fig. 11 shows the two different output buffers for a dedicated output network for a DPU output sending data to two receivers (Load1 and Load2). Fig. 11(b) also shows the gating locations and control signals (en1 and en2). We call such output buffers split output buffers. For the shared output network implemented in the trunk-branches style, there are shared and dedicated parts in the output network. To maximize the benefit of SSA gating, instead of using a single large output buffer, we need to use split buffers, $i.e.$, a buffer for the shared part and a dedicated buffer for each dedicated part. Then we can gate SSA before the dedicated buffer. Fig. 12 shows the two different output buffers for a trunk-branches output network. It only shows two receivers (Load1 and Load2). Fig. 12(b) also shows the gating locations and control signals (en1 and en2). In both the dedicated and shared output network cases, split buffers consume no more power or area than the single large buffer. Meanwhile, split buffers facilitate SSA gating.

Figure 11: Output buffers for a fully dedicated output network: (a) single large buffer, and (b) split buffers with signal gating

Figure 12: Output buffers for a trunk-branches output network: (a) single large buffer, and (b) split buffers with signal gating.