Technical note · Aug 29, 2026 · 8 min read

How Much of a Small Chip Actually Computes?

Physical-implementation overhead in tiny arithmetic blocks, measured on an open 130nm flow

Subomi Olagoke

Abstract

A synthesized netlist is logic. A layout that a foundry could make also carries the machinery that lets the logic survive manufacture: antenna diodes that bleed off etch charge before it destroys a gate, well taps that keep the substrate from latching up, buffers inserted to close timing and fix hold, and fill that pads the die to a target density. Writeups usually fold this overhead into a single area number, and I have not found it tracked as a block scales. I took one operation, a signed int8 multiply-accumulate, through the open SkyWater 130nm flow at lane counts from 1 to 64, split every placed layout into logic and infrastructure, and followed the split as the block grew. The arithmetic share neither holds nor improves with size. It falls from 59% of the placed cells at one lane to 48% at sixty-four, dropping below half on the way, so a large multiply-accumulate is mostly made of something other than multiply-accumulate. The overhead outgrows the logic instead of amortizing, and its parts scale against different quantities: well taps against die area, antenna diodes against routed wirelength. The wirelength itself grows faster than the cell count, and that is what erodes the logic share.

1Why measure this

Almost every writeup describes a hardware block with two numbers, its area and its clock. Both come out of physical implementation, and both hide what physical implementation did. Between a gate-level netlist and a GDSII the tools add a surprising amount of matter that is not logic, for physical reasons unrelated to what the block computes. A reader who is told the block is, say, a tenth of a square millimetre will tend to assume that tenth is arithmetic, and it is not.

At any single design point the gap is easy to see. In an earlier note [1] I took this same multiply-accumulate to a layout and found that of roughly 7,600 placed standard cells, only about half did any arithmetic. The rest were 1,759 antenna diodes, 1,440 well taps and a few hundred buffers, with thousands of fill cells beyond that. One point is an anecdote. This note asks whether it is a curve: as the block grows, does the overhead shrink as a share, hold constant, or grow, and do its parts move together or separately?

2The block and the flow

The device under test is mak, a signed int8 multiply-accumulate lane. It multiplies LANES activation and weight bytes, sums the products through an adder tree, and accumulates the result across cycles into a fixed 32-bit register. It is the atomic operation of a small quantized matmul, and LANES is its only parameter and the only thing swept here. The accumulator stays 32 bits wide throughout, so the sequential state is held constant on purpose while the combinational arithmetic grows.

Every point uses the same flow and the same targets. The PDK is the open SkyWater 130nm sky130A [5], synthesis is Yosys [4], and placement and routing are OpenROAD [3] under OpenLane 2.3.10 [2], at a fixed 40 ns clock and a fixed 40% core utilization target, with heuristic antenna-diode insertion enabled. That mode (RUN_HEURISTIC_DIODE_INSERTION) places a diode on any net whose estimated Manhattan length at global placement exceeds a PDK threshold (HEURISTIC_ANTENNA_THRESHOLD, 90 here), whether or not the net would violate the antenna rule, and antenna repair during global routing (GRT_REPAIR_ANTENNAS) adds more where violations remain. The heuristic can insert more diodes than are strictly needed, so the diode counts below are counts under that setting and should not be read as a minimum. Only LANES changes, across 1, 2, 4, 8, 16, 32 and 64. Each run is signed off with design-rule checks (Magic [6] and KLayout [8]) and layout-versus-schematic checks (Netgen [7]), and each emits a machine-readable metrics file. The cell census, wirelength, die area and worst-corner timing slack are all read from that file; nothing below is hand-counted.

3Results

For each lane count the placed layout is split into logic (the combinational arithmetic, its inverters, and the 32 accumulator flip-flops) and the machinery around it: antenna diodes, well taps, and timing and hold buffers. Die area, routed wirelength and worst-corner setup slack are recorded alongside. All seven runs pass the design-rule and layout-versus-schematic checks.

LANES logic diodes taps buffers total cells logic % (count) logic % (area) die (µm²) wire (mm)
1 619 80 220 128 1,054 58.7% 77.9% 20,862 13.9
2 1,076 207 384 163 1,838 58.5% 78.8% 34,222 23.4
4 2,005 792 731 247 3,783 53.0% 76.9% 61,190 48.1
8 3,979 1,710 1,440 434 7,571 52.6% 77.9% 115,201 105.1
16 7,859 3,427 2,856 774 14,924 52.7% 79.6% 219,286 232.0
32 16,400 8,974 6,050 1,511 32,942 49.8% 78.0% 448,891 543.3
64 31,827 20,073 11,795 3,023 66,726 47.7% 77.2% 865,429 1,276.4

The total also includes the 7 or 8 clock-tree buffers and inverters at each size, which fall in none of the other columns. The buffers column counts the flow's timing-repair buffers, hold-fixing delay cells included. Every share in this note is a share of placed standard cells by count unless marked as area. Counting gives a diode or a tap (1.25 µm² each in this library) the same weight as a logic cell (7.8 µm² on average at eight lanes), and by cell area the picture changes. Logic holds between 77% and 80% of cell area at every lane count, antenna diodes grow from 3% of it at one lane to 14% at sixty-four, and taps stay near 4 to 5%. The area shares come from the final placed DEF of each run and the cell sizes in the sky130_fd_sc_hd LEF, and they sum to the flow's reported standard-cell area.

Layout composition by lane count
FIG ALayout composition by lane count
Figure 1. The composition of the placed layout as a share of standard cells, by lane count. The black band is logic. It starts above half and slides beneath the fifty-percent line as the antenna diodes (in pink) take a growing share.
Diodes track wiring, taps track area
FIG BDiodes track wiring, taps track area
Figure 2. Logic, antenna diodes, well taps and routed wirelength against lane count, log-log. The diode curve runs parallel to the wire curve, and both are steeper than the logic curve.

Three results come out of the table.

  1. The arithmetic share falls and drops below half. Logic is 58.7% of the placed cells at one lane and 47.7% at sixty-four, passing 50% somewhere between sixteen and thirty-two lanes (Figure 1). A bigger block does not mean a larger logic share. The overhead behaves less like a fixed tax that a larger block dilutes and more like a cost that grows faster than the arithmetic it exists to protect.
  2. The overhead's parts scale against different quantities, and only one of them is logic. Well taps stay at roughly a fifth of the cells across the whole range, because they sit on a fixed placement grid and so follow die area instead of gate count. Timing buffers stay small. The accumulator is thirty-two flip-flops at every size, a constant that is three percent of the cells at one lane and rounding error by sixty-four. The antenna diodes are what moves. From four lanes up they hold at fifteen to seventeen per millimetre of routed wire (about six and nine per millimetre at one and two lanes), so they follow the wiring (Figure 2), and their share of the cells climbs from eight percent to thirty.
  3. Wiring drives the decline. Routed length grows super-linearly, from 14 mm at one lane to 1.28 metres at sixty-four, faster than the cell count, so the metal spent per logic cell rises from 22 to 40 µm as the datapath widens and its nets reach further. Since the diodes follow that wire, they outpace the logic, and the logic share erodes. What moves the whole result is how much metal it takes to connect the arithmetic.

4Discussion

All of this bears on how a small hardware result should be read and reported. "A 0.1 mm² MAC" invites the assumption that the MAC is 0.1 mm² of MAC. At small sizes that is far off, and the difference is structural and predictable. The infrastructure is neither waste nor optional, since a layout without the diodes fails fabrication and one without the taps latches up. It is also not what the block is for, and folding it into a single area figure overstates how much silicon the arithmetic costs.

There is a practical corollary for anyone estimating an accelerator from a tile. An accelerator is thousands of these, so the per-tile overhead is paid thousands of times, and the direction the fraction moves is a design lever worth attention. The comfortable assumption is that a wider tile amortizes the overhead. The data point the other way. A wider tile buys a worse logic fraction, because the wiring, and the antenna diodes that follow it, grow faster than the multipliers do. On this axis the overhead is paid many times over and also takes a growing share of each tile as the tile widens, which argues for many small tiles over a few large ones. That conclusion rests on one process and one block shape, but it runs against the usual intuition, and that intuition tends to go unstated in a napkin estimate.

5Limitations

This is one process, one flow, one block shape, and open-flow defaults. A different PDK, a different utilization target, or a commercial flow with more aggressive optimization would move the absolute numbers, and possibly the trend. The block is combinational-heavy with a single small register, so the constant sequential cost is unusually visible; a register-heavy block would tell a different story. The antenna and tap counts are artifacts of a particular diode-insertion heuristic and a particular tap spacing, both configurable. Nothing here is fabricated: these are signed-off layouts, not measured silicon. The claim is narrow. On this flow and at these targets, the composition of a small MAC is what the numbers above say, and it changes with size.

6Conclusion

The headline number on a small hardware block, its area, mostly measures something other than what readers assume. A large share of a tiny layout is machinery for surviving fabrication, that share moves predictably with size, and its parts scale separately against logic, wiring and area. Reporting the split alongside the total costs nothing, and it is more accurate than letting one area figure imply that all of it computes.

References

  1. S. Olagoke. mak, case study, 2026. https://www.subomi.net/mak
  2. Efabless. OpenLane 2 (version 2.3.10). https://github.com/efabless/openlane2
  3. T. Ajayi et al. Toward an open-source digital flow: first learnings from the OpenROAD project. Proceedings of the 56th Design Automation Conference (DAC), 2019. https://doi.org/10.1145/3316781.3326334
  4. C. Wolf and J. Glaser. Yosys: a free Verilog synthesis suite. Proceedings of the 21st Austrian Workshop on Microelectronics (Austrochip), 2013. https://yosyshq.net/yosys/files/yosys-austrochip2013.pdf
  5. Google and SkyWater Technology. SkyWater SKY130 open-source PDK. https://github.com/google/skywater-pdk
  6. Magic VLSI Layout Tool. Open Circuit Design. http://opencircuitdesign.com/magic/
  7. Netgen. Open Circuit Design. http://opencircuitdesign.com/netgen/
  8. KLayout. https://www.klayout.de/