22nm CMOS Datapath & Memory

Project 1: 8-bit Ripple-Carry Adder

An 8-bit ripple-carry adder, built and simulated in Electric VLSI. The full adder's outputs are:

Both depend on A ⊕ B, which made XOR2 the natural primitive to build around. Compute it once, feed it into both the sum and carry paths. Several alternative sum and carry topologies were explored and rejected along the way; that analysis is in the full report. Once the topology was settled, the interesting question became which XOR2 implementation to use.

Two candidates were carried forward and compared head-to-head at the full 8-bit system level:

Baseline: four-NAND2 XOR2 (16 T). Every node fully restored, easy to analyze, conservative.

Optimization: transmission-gate XOR2 with two-inverter output buffer (10 T). Same logic, 40% fewer transistors, and a shorter signal path. The buffer restores the TG's voltage output to a full rail-to-rail logic level before the next stage.

Unfortunately the only screenshot of the full adder I had saved is cropped a little.

The full adder bit-slice combines the two primitives. One NAND2 computes AB in parallel with the first XOR2 (P = A ⊕ B). That intermediate P feeds both the sum path (a second XOR2 producing S = P ⊕ C_in) and the carry path, where two right-hand NAND2 cells realize C_out = AB + C_in(A ⊕ B).

Sharing P between sum and carry is what makes this topology efficient: the most expensive intermediate is computed once and reused. Eight of these bit-slices (plus a half adder at bit 0) make up the full 8-bit ripple-carry adder.

SPICE Results

Side-by-side SPICE waveforms. Left column: optimized TG-XOR design. Right column: NAND4-XOR baseline. The bottom row is full carry propagation (255 + 1), the worst case. Optimized settles 44 ps earlier.

The Leakage Tradeoff

The optimized design wins on delay and area, but transmission gates create a partially-conducting path from supply to ground at the worst-case input (A = B = 1). That's a ~10× leakage penalty vs. the fully static NAND-only baseline. At minimum-leakage inputs the two designs are comparable. The choice between them depends on duty cycle: high-throughput, mostly-switching paths want the optimized cell; always-on, low-activity datapaths want the baseline.

Project 2: 16×4 SRAM

A synchronous 16×4 SRAM (16 words of 4 bits) in the same 22nm process, targeting at least 500 MHz operation on a single clock input. The design spans the full memory system: a 6T bit cell, non-overlapping clock generation, row and column decoders, sense amplifiers, column drivers, and a voltage midpoint reference. It is graded on a figure of merit that combines bitcell area, power, and delay squared, so every sub-block has to be sized for both correctness and score.

Top-Level Architecture

The full memory fits on a single schematic. Address inputs A₀–A₃ are latched and fed through a 4-to-16 row decoder gated by φ₂, driving sixteen word-lines into the 16×4 bitcell array. Four column drivers precharge BL/BLb during φ₁ and drive writes during φ₂; four isolated latch-type sense amps are armed by SAE = delayed(φ₂·Wr) and resolve reads onto a per-column bus shared with the write path. A two-phase non-overlapping clock generator takes the single external Clk and produces the two phases that keep precharge and access strictly separated.

Top-level schematic. All control signals (PCHb, SAE, WrEn, the inverted Wr) come from the two-phase clock and the Wr fanout network along the bottom of the diagram.

Two design decisions drive the rest of the implementation. First, PCHb is wired directly to φ₁, which forces precharge to finish before φ₂ can rise. The non-overlap gap is what guarantees the column drivers and word-lines never fight each other. Second, the sense amp is an isolated latch-type pair armed by a φ₂-derived SAE, so it only fires once the bitline differential has developed past its offset. Every other block exists to feed these two.

One Full Clock Cycle

The easiest way to see the whole design work at once is to overlay every relevant signal on one clock period. I ran characterization at 320 ps (3.125 GHz), which is a deliberately safe choice: about 7% above the ~298 ps absolute minimum found by binary search further down, so no waveform is sitting right on its own failure boundary.

One 320 ps cycle, split into four phases. During precharge (roughly 0–150 ps), PCHb is low and BL/BLb are pulled back up to V_DD. A short non-overlap gap separates precharge from access. When φ₂ rises, WL fires, the bitlines start to develop a differential, and a delayed SAE arms the sense amp only after that differential is past 50 mV. Q/D resolves cleanly well before the next precharge begins.

Functional Verification

Correctness was validated bottom-up, from the bitcell up to the full array. The most interesting test is mainTest2: two different 4-bit words written to two different addresses, then read back. This checks that the decoder routes the correct row and that storing one word does not disturb another.

mainTest2: 1001 → address 0000, then 0110 → address 1111, then read both back. Each read matches the written word exactly; no cross-talk between rows.

How Fast Can It Go?

Minimum operating period was found with a binary search: shrink the clock period until reads stop matching the values that were written, then bracket the break point.

Binary search for T_min. The read-back voltage on a known-0 cycle sits cleanly at 0 V for long periods and snaps up to V_DD for short ones; the transition crosses the V_DD/2 decision threshold near 298 ps, setting f_max ≈ 3.36 GHz, well above the project's 500 MHz requirement. The 320 ps operating point used for the earlier waveforms and for the power measurement sits about 7% above this break point, so the reported numbers are not riding the edge of correctness.

Summary of Metrics