← Back to the chapter · Previous detail · Next detail →

Technical detail · Energy Cost of Common Operations

Arithmetic and memory at 45 nm

OperationEnergy per operationIllustrative circuit area
8-bit integer add0.03 pJ36 μm2\mu\text{m}^2
16-bit integer add0.05 pJ67 μm2\mu\text{m}^2
32-bit integer add0.10 pJ137 μm2\mu\text{m}^2
16-bit floating-point add0.40 pJ1,360 μm2\mu\text{m}^2
32-bit floating-point add0.90 pJ4,184 μm2\mu\text{m}^2
8-bit integer multiply0.20 pJ282 μm2\mu\text{m}^2
32-bit integer multiply3.1 pJ3,495 μm2\mu\text{m}^2
16-bit floating-point multiply1.1 pJ1,640 μm2\mu\text{m}^2
32-bit floating-point multiply3.7 pJ7,700 μm2\mu\text{m}^2
32-bit read from an 8 kB static RAM (SRAM)5 pJMemory-dependent
32-bit DRAM read640 pJOff-core memory-dependent

Walk down the table rather than reading across it. A 32-bit SRAM read costs about 50 times the listed 32-bit integer add. The DRAM read costs about 6,400 adds, or about 173 of the listed 32-bit floating-point multiplies. The hierarchy is therefore not simply “integer cheap, floating point expensive.” Once operands travel to distant memory, movement can dominate either arithmetic choice.

← Back to the chapter · Previous detail · Next detail →