Four register banks hang off a three-level clock tree. Decide where clock gates go, low at the banks or high in the tree, and meet a power budget.
Hardware context: Register banks under a clock tree, and a bank of transistor-level flops. Place clock gates high in the tree and share local clock drivers within a power budget.
A block has four register banks on one 10 ns clock tree. The root buffer feeds two branch buffers. Branch 0 feeds banks a (24 flops) and b (8 flops). Branch 1 feeds banks c (16 flops) and d (4 flops). Each bank has its own leaf buffer.
The block is busy in 30 % of the cycles and idle in the rest. Bank b loads in every busy cycle. Bank a loads in 5 % of all cycles. Bank c loads in 2 % and bank d in 3 %, never in the same cycle.
Today every bank holds its value with enable muxes, and the block burns 404 µW. The budget is 55.5 µW.
Choose a build for each bank: enable muxes, or a clock gate behind its leaf buffer. You may also put a gate in front of a branch buffer or behind the root buffer. You choose the signal that enables it. A gate higher in the tree must stay open whenever any bank below it loads.
Requirements
| Requirement | Limit | Unit |
|---|---|---|
| Gates whose enable does not cover every bank below them | = 0 | count |
| Clock-gating efficiency (share of flop-cycles with the clock stopped) | ≥ 0.9 | count |
Budget: Total power of the block ≤ 55.5 µW. Smaller is better after every requirement passes.
Recorded experiments: Recorded build-time OpenSTA 3.1.0 report_power on sky130_fd_sc_hd at tt, 25 °C, 1.8 V, with a 10 ns clock. Data inputs toggle 0.1 times per cycle, and each enable has its stated duty. The clock tree is drawn in the netlist, with ideal wires (no placement). No tool runs during grading. Circuit: a clkbuf_8 root, two clkbuf_8 branch buffers, a clkbuf_4 leaf buffer per bank and dfxtp_1 flops. A bank uses a mux2_1 per flop or one dlclkp_1 behind its leaf buffer. A branch gate is a dlclkp_2 in front of the branch buffer; the root gate is a dlclkp_4 behind the root buffer. The probe measured all 128 builds.
Full specification
| JSON field | Allowed values | Meaning |
|---|---|---|
| bank_a, bank_b, bank_c, bank_d | "mux", "icg" | Enable muxes in front of every flop, or one clock gate (dlclkp_1) behind the leaf buffer of the bank. |
| root | "none", "en_blk" | A clock gate (dlclkp_4) between the root buffer and both branch buffers. |
| branch0 | "none", "en_br0", "en_a", "en_b" | A clock gate (dlclkp_2) in front of branch buffer 0, and its enable. |
| branch1 | "none", "en_br1", "en_c", "en_d" | A clock gate (dlclkp_2) in front of branch buffer 1, and its enable. |
| Enable | Share of cycles at 1 |
|---|---|
| en_blk | 30 % |
| en_br0 | 30 % |
| en_br1 | 5 % |
| en_a | 5 % |
| en_b | 30 % |
| en_c | 2 % |
| en_d | 3 % |
A gate in front of a buffer stops that buffer and everything below it. A gate behind a leaf buffer stops only the flops of its bank.
Clock-gating efficiency is the share of flop-cycles in which a flop gets no clock edge. It counts flops only, not buffers.
Recorded build-time OpenSTA 3.1.0 report_power on sky130_fd_sc_hd at tt, 25 °C, 1.8 V, with a 10 ns clock. Data inputs toggle 0.1 times per cycle, and each enable has its stated duty. The clock tree is drawn in the netlist, with ideal wires (no placement). No tool runs during grading. Circuit: a clkbuf_8 root, two clkbuf_8 branch buffers, a clkbuf_4 leaf buffer per bank and dfxtp_1 flops. A bank uses a mux2_1 per flop or one dlclkp_1 behind its leaf buffer. A branch gate is a dlclkp_2 in front of the branch buffer; the root gate is a dlclkp_4 behind the root buffer. The probe measured all 128 builds.
Numeric checks use inclusive bounds. Only the documented fields are accepted.
Examples
Example 1
Enable Muxes Everywhere- Input
{ "bank_a": "mux", "bank_b": "mux", "bank_c": "mux", "bank_d": "mux", "root": "none", "branch0": "none", "branch1": "none" }- Output
Gates whose enable does not cover every bank below them: 0 Clock-gating efficiency (share of flop-cycles with the clock stopped): 0 Total power of the block: 404 µW
Every flop and every buffer sees every clock edge.
Example 2
A Clock Gate on Every Bank- Input
{ "bank_a": "icg", "bank_b": "icg", "bank_c": "icg", "bank_d": "icg", "root": "none", "branch0": "none", "branch1": "none" }- Output
Gates whose enable does not cover every bank below them: 0 Clock-gating efficiency (share of flop-cycles with the clock stopped): 0.922 Total power of the block: 95.15 µW
The efficiency is high, yet every buffer still toggles in every cycle.
Example 3
Every Gate the Tree Can Hold- Input
{ "bank_a": "icg", "bank_b": "icg", "bank_c": "icg", "bank_d": "icg", "root": "en_blk", "branch0": "en_br0", "branch1": "en_br1" }- Output
Gates whose enable does not cover every bank below them: 0 Clock-gating efficiency (share of flop-cycles with the clock stopped): 0.922 Total power of the block: 56.48 µW
Some of these gates never close on their own.
Submission Rules
- Submit one JSON object with every documented field.
- Run checks your choice against the recorded measurements. Submit applies the same rules and records the result.
- The numbers are estimates for one library corner, one set of duties and ideal wires. Passing does not sign off the power of a chip.
Hints
Hint 1
Compare the clock power of the leaves-only build with its efficiency. Which cells does the efficiency not count?
Hint 2
For each high gate, ask how often its enable is 0 while the gate above it is open.
Hint 3
A bank gate whose enable equals the enable of a gate above it closes only when that gate is already closed.
What is checked: Recorded build-time OpenSTA 3.1.0 report_power on sky130_fd_sc_hd at tt, 25 °C, 1.8 V, with a 10 ns clock. Data inputs toggle 0.1 times per cycle, and each enable has its stated duty. The clock tree is drawn in the netlist, with ideal wires (no placement). No tool runs during grading. Circuit: a clkbuf_8 root, two clkbuf_8 branch buffers, a clkbuf_4 leaf buffer per bank and dfxtp_1 flops. A bank uses a mux2_1 per flop or one dlclkp_1 behind its leaf buffer. A branch gate is a dlclkp_2 in front of the branch buffer; the root gate is a dlclkp_4 behind the root buffer. The probe measured all 128 builds.
Run and Submit use the same measured requirements. Run is a preview; Submit records an attempt. No physical tool runs during this exercise; the records come from the pinned build-time experiments.
More problems on this topic