Random-access memory

Memory manufacturing depends partly on the silicon area required for the installed capacity. With the historical model and an assumed density of 0.625 GB/cm², a 32 GB module gives 117.86 kg CO₂ eq.; 5.22 kg is only the fixed term. HBM packaging multipliers used in the simulator are exploratory assumptions, not validated universal ratios.

Manufacturing contributions

DRAM fabrication involves cleanrooms, lithography, etching and metallisation; see Semiconductors. Capacity depends on cell density, die area, die count and module organisation. HBM stacks DRAM dies with through-silicon interconnections; this differs from the memory-cell layers of 3D NAND.

These additional processes require their own inventories. Their existence alone does not establish a numerical impact multiplier.

Parametric model

Embodied formula

$$I_{emb}=\left(\frac{C}{D}C_{die}+C_{base}\right)F_{pkg}N_{modules}$$

$C$ is capacity per module in GB; $D$ is density in GB/cm²; $C_{die}$ is impact per cm²; $C_{base}$ is the fixed module contribution; $F_{pkg}$ is an assumed packaging multiplier; $N_{modules}$ is module count.

Historical coefficients

Indicator	Die coefficient per cm²	Base per module	Impact unit
GWP	2.20	5.22	kg CO₂ eq.
ADP	6.30E-05	1.69E-03	kg Sb eq.
Primary energy	27.30	74.00	MJ

The historical model is documented in Boavizta's RAM methodology, based on Green Cloud Computing. Combining ADP with another dataset's ADPe requires compatible characterisation methods.¹²

Memory families

Family	Application	Configuration considerations
DDR4 / LPDDR4	System or mobile memory	Packaging and module organisation differ; LPDDR devices are not necessarily DIMMs
DDR5 / LPDDR5	System or mobile memory	Density, generation and power management require specific data
GDDR6 / GDDR6X	Graphics memory	Board-mounted memory and bandwidth-oriented design
HBM2e	Some datacentre accelerators, including A100 variants	Stacked dies and package integration
HBM3 / HBM3e	H100 uses HBM3; H200/B200 use HBM3e	Capacity and stack configuration depend on the exact product

Density and packaging scenarios

These are D4B exploratory hypotheses, not guaranteed generation densities or validated packaging factors. Their applicability depends on die count, surface, counted capacity and the boundary between memory, interposer and accelerator.

Type	Assumed density, GB/cm²	Assumed packaging multiplier
DDR4 / LPDDR4	0.625	1.0
DDR5 / LPDDR5	1.0	1.1
GDDR6	0.8	1.3
HBM2e	0.5	3.0
HBM3	0.8	3.5

GDDR's integration and bandwidth requirements do not alone establish a higher impact per GB than DDR. A comparison needs compatible inventories.

Reproducible scenarios

Configuration	Modelled embodied GWP	Historical ADP result	Interpretation
One 32 GB DDR4 module	117.86 kg CO₂ eq.	4.92E-03 kg Sb eq.	Historical base model with assumed density
Sixteen 32 GB DDR5 modules	1,330.91 kg CO₂ eq.	6.52E-02 kg Sb eq.	Includes the hypothetical 1.1 multiplier
Six 16 GB HBM2e units	1,361.16 kg CO₂ eq.	Not reported	Hypothetical configuration, not an H100 inventory

The latter calculations are $(32/1\times2.2+5.22)\times1.1\times16$ and $(16/0.5\times2.2+5.22)\times3\times6$. Arithmetic consistency does not validate the multipliers or a universal RAM share in server manufacturing.

Operational model

A simplified constant-capacity model is:

$$P_{ram}=C_{total}\times p$$

Historical examples use $p=0.392$ W/GB in Cloud Carbon Footprint or $p=0.284$ W/GB in the Boavizta/UBA model.³

Capacity	At 0.392 W/GB	At 0.284 W/GB
16 GB	6.3 W	4.5 W
128 GB	50 W	36 W
512 GB	201 W	145 W
2,048 GB	803 W	582 W

Refresh and data retention consume energy, but accesses, frequency and power states also matter. A constant W/GB coefficient is not a measured load-response curve or a universal law across DDR, LPDDR and HBM.

Indicator coverage

Indicator	Treatment
GWP	Historical area model plus explicit scenario assumptions
ADP / ADPe	Historical coefficients; method compatibility required
Primary energy	Historical coefficients, distinct from electrical kWh
WU	No validated characterised factor here; removed from simulator results
AP, PM, toxicity	Whole-equipment factors may include them without isolating memory
IR	Compatible manufacturing/use inventories and LCIA units required

A physical water quantity per die area cannot be added directly to characterised electricity-related water. Missing component results do not imply zero impact. The current interactive output focuses on GWP and electricity rather than claiming all these indicators are calculated.

Assumptions and limitations

Capacity alone is insufficient to establish a modern memory inventory. HBM multipliers require calibration; integrated memory and system RAM must have distinct boundaries. Power coefficients vary in suitability by technology and workload. A complete accelerator inventory that includes HBM already covers that contribution.

Interactive example

Memory type

Capacity per module (GB)

Number of modules

Lifetime (years)

Daily usage (hours)

Country

PUE

Results (per year)

Climate (GWP)

Embodied: 23.57 kgCO2eOperational: 5.55 kgCO2eTotal: 29.12 kgCO2e

Embodied

Operational

Operational electricity

173.74 kWh

Parametric model based on Boavizta (2.20 × die area in cm² + 5.22 kgCO₂e per module). Power: 0.392 W/GB (Cloud Jewels). Densities and HBM multipliers are uncalibrated exploratory assumptions.

Use-phase electricity: D4B snapshot, 2025 annual averages. The selected country applies to use; it does not relocate manufacturing factors.

Footnotes

1. Green Cloud Computing, UBA, 2021 ; coefficients par cm² de die DRAM. Implémentés dans Boavizta RAM methodology. ↩

2. Versioned cloud CPU/accelerator records, identified memory configurations and historical Boavizta component models. Digital4Better data catalogue, snapshot 77cae4a. References and scope: Sources and corpus. ↩

3. Cloud Carbon Footprint methodology. Cloud Carbon Footprint methodology. Online documentation, unpinned version. ↩