system-energy

System-level energy comparison across post-CMOS technologies. Applies control, memory, and I/O overhead to device-level per-MAC energies, producing the honest wall-plug comparison that raw device papers skip.

The result

128x128 matrix-vector multiply (16,384 MACs):

Among measured technologies, adiabatic CMOS wins by 6.8x to 481x.

The model

system_energy = wall_energy × (1 + control_mult + memory_mult)

Where:

• wall_energy = device energy × cooling multiplier (1 for room-temp, 1500 for 4K)
• control_mult = clock distribution, level shifters, drivers
• memory_mult = SRAM read/write per MAC

Overhead factors (from IRDS 2023 and vendor datasheets):

Caveat: these multipliers are engineering estimates, not rigorous derivations. They reflect typical proportions in published system analyses. The ordering is robust to ±50% changes in any single multiplier.

Why system-level matters

Device-level papers compare switches to switches:

• SFQ: 0.207 aJ junction vs 245 aJ CMOS switch
• Photonic: 5 fJ per MAC vs 245 aJ CMOS switch

These comparisons are misleading because they ignore:

1. Cooling: SFQ's 0.207 aJ junction becomes 310 aJ wall
2. Transceivers: photonic's 1 fJ optical compute becomes 5,000 aJ with modulator + detector + laser
3. Peripherals: memristor's 1.5 fJ device becomes 6,000 aJ with row/column decoders and ADCs

The system-level view is where the honest verdict lives.

Install

pip install -e .

CLI

system-energy table                    # comparison table
system-energy table --n_macs 78400     # MNIST-scale task
system-energy plot --output verdict.png

Python API

from system_energy import TECHNOLOGIES, system_energy, compare_all

# Single technology
E = system_energy("memristor")     # 6000.0 (aJ per MAC)

# All
all_e = compare_all()

Tests

pytest tests/ -v

License

Apache 2.0