A 6-bit register bank burns too much clock energy. Rebuild it from flop cores that share one local clock driver, and keep every bit correct at 1 GHz.
Circuit 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.
Six dfxtp_1 flops make up a 6-bit register bank. Each flop has its own local clock driver: two inverters that make clkneg and clkpos from CLK. The bank runs at 1 GHz.
The library has no multi-bit flop, but you may build one. The core of the flop is the 20 transistors that only read clkneg and clkpos. Drive several cores from one driver.
Rebuild the subcircuit bank6. Its energy per cycle, with the last clock-tree buffer, must be at most 190 fJ. Every bit must still capture its data at 1 GHz.
What the Grader Measures
Your subcircuit bank6 (ports clk, d0, d1, d2, d3, d4, d5, q0, q1, q2, q3, q4, q5, vdd) under the grader's own bench, one run per corner (tt):
| Check | Corner | Rule |
|---|---|---|
| Bits captured wrong | tt | ≤ 0 |
| Energy per cycle | tt | ≤ 190 fJ |
Full specification
.subckt bank6 clk d0 d1 d2 d3 d4 d5 q0 q1 q2 q3 q4 q5 vdd
* ffcore, sky130_fd_pr__nfet_01v8 and sky130_fd_pr__pfet_01v8_hvt instances only
.ends
.subckt ffcore d ckn ckp q vdd
* master and slave latches of dfxtp_1: capture d on the rising edge of clk,
* where ckn = not clk and ckp = clk, both made by a local driver
.ends
| Check | Rule | How the grader measures it |
|---|---|---|
| Wrong bits | = 0 | Each q is read just before the next rising edge, in six cycles for each of the six bits: 36 samples |
| Energy per cycle | ≤ 190 fJ | Charge from the bank supply and from the last tree buffer's supply over five cycles, times 1.8 V, divided by five |
The grader runs once at tt, 27 °C. The clock comes from one inverter (W_N = 1 µm, W_P = 2 µm) and has 1 ns periods. Each data input changes 0.3 ns after a rising edge.
The SKY130 models start at a width of 0.42 µm, so the core simulates its four 0.36 µm transistors at 0.42 µm.
Examples
Example 1
Six Private Drivers, as Shipped- Design
.subckt bank6 clk d0 d1 d2 d3 d4 d5 q0 q1 q2 q3 q4 q5 vdd XP0_1n n0 clk 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP0_1p n0 clk vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XP0_2n p0 n0 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP0_2p p0 n0 vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XC0 d0 n0 p0 q0 vdd ffcore XP1_1n n1 clk 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP1_1p n1 clk vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XP1_2n p1 n1 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP1_2p p1 n1 vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XC1 d1 n1 p1 q1 vdd ffcore XP2_1n n2 clk 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP2_1p n2 clk vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XP2_2n p2 n2 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP2_2p p2 n2 vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XC2 d2 n2 p2 q2 vdd ffcore XP3_1n n3 clk 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP3_1p n3 clk vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XP3_2n p3 n3 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP3_2p p3 n3 vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XC3 d3 n3 p3 q3 vdd ffcore XP4_1n n4 clk 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP4_1p n4 clk vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XP4_2n p4 n4 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP4_2p p4 n4 vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XC4 d4 n4 p4 q4 vdd ffcore XP5_1n n5 clk 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP5_1p n5 clk vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XP5_2n p5 n5 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.15 XP5_2p p5 n5 vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.15 XC5 d5 n5 p5 q5 vdd ffcore .ends
- Measured by the grader
Bits captured wrong: 0 (needs ≤ 0) Energy per cycle: 224.1 fJ (needs ≤ 190 fJ)
Every bit is right, and the clock costs too much.
Example 2
One Shared Driver with Long Transistors (L = 0.4 µm)- Design
.subckt bank6 clk d0 d1 d2 d3 d4 d5 q0 q1 q2 q3 q4 q5 vdd XS1n ckn clk 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.4 XS1p ckn clk vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.4 XS2n ckp ckn 0 0 sky130_fd_pr__nfet_01v8 W=0.42 L=0.4 XS2p ckp ckn vdd vdd sky130_fd_pr__pfet_01v8_hvt W=0.64 L=0.4 XC0 d0 ckn ckp q0 vdd ffcore XC1 d1 ckn ckp q1 vdd ffcore XC2 d2 ckn ckp q2 vdd ffcore XC3 d3 ckn ckp q3 vdd ffcore XC4 d4 ckn ckp q4 vdd ffcore XC5 d5 ckn ckp q5 vdd ffcore .ends
- Measured by the grader
Bits captured wrong: 14 (needs ≤ 0) Energy per cycle: 99.6 fJ (needs ≤ 190 fJ)
Cheap and wrong: its edges are too slow for 1 GHz.
Example 3
One Shared Driver Four Times the Cell Size- Design
.subckt bank6 clk d0 d1 d2 d3 d4 d5 q0 q1 q2 q3 q4 q5 vdd XS1n ckn clk 0 0 sky130_fd_pr__nfet_01v8 W=1.68 L=0.15 XS1p ckn clk vdd vdd sky130_fd_pr__pfet_01v8_hvt W=2.56 L=0.15 XS2n ckp ckn 0 0 sky130_fd_pr__nfet_01v8 W=1.68 L=0.15 XS2p ckp ckn vdd vdd sky130_fd_pr__pfet_01v8_hvt W=2.56 L=0.15 XC0 d0 ckn ckp q0 vdd ffcore XC1 d1 ckn ckp q1 vdd ffcore XC2 d2 ckn ckp q2 vdd ffcore XC3 d3 ckn ckp q3 vdd ffcore XC4 d4 ckn ckp q4 vdd ffcore XC5 d5 ckn ckp q5 vdd ffcore .ends
- Measured by the grader
Bits captured wrong: 0 (needs ≤ 0) Energy per cycle: 199.2 fJ (needs ≤ 190 fJ)
Correct, and 199 fJ: a big driver spends the saving on its own capacitance and misses the budget.
Constraints and Contract
- The netlist defines .subckt bank6 exactly once, with the ports shown above in that order.
- Inside bank6: X instances of ffcore, sky130_fd_pr__nfet_01v8 or sky130_fd_pr__pfet_01v8_hvt, with comments and .param lines. No sources, resistors or capacitors.
- The grader writes the .lib line itself; everything outside bank6 is replaced by its bench.
Hints
Hint 1
Run the starter. Then count how many inverter inputs the clock-tree buffer drives.
Hint 2
One driver can serve every core. Keep ckn the inverse of clk and ckp the inverse of ckn.
Hint 3
A weak driver makes slow ckn and ckp edges. The latches then misbehave at speed, and slow edges also waste energy in the cores.
What the grader checks: The grader places your bank6 under its own bench at tt. The bench has a 1 GHz clock from one inverter, six data patterns and 36 samples. It runs on ngspice 42 with the pinned SKY130 models. One corner and one speed are covered; wires, hold margins at other corners and layout are not.
Run simulates your netlist as written, with your own analyses and measurements. Submit sends it to the grader, which keeps only the subcircuit bank6 and makes 1 ngspice run. The numbers come from ngspice and the SKY130 models the site pins; they prove the measured cases, not every condition a chip meets.
See all 2 steps and the final capstone