HBM silicon intensity and allocation

The mechanism

HBM requires silicon for memory arrays, peripheral circuits, vertical connections and base-die functions, followed by thinning, stacking and test. A stack’s footprint hides the area of its multiple dies. Silicon intensity can mean geometric area per bit or empirical wafer supply per delivered bit: different accounting methods. The geometric route explicitly applies relevant yields. An empirical ratio may already include those losses and must not receive them again. This lesson converts memory requirements into a consistent supply boundary with a dated supplier comparison whose denominator is kept explicit.

Specify gross, good-die, tested-stack or package-delivered bits as the output boundary.
For geometry, sum die area and apply explicitly defined, nonoverlapping yield losses.
For empirical intensity, determine which wafer-to-output losses the reported ratio already includes.
Calculate separate memory and base-die process demands without applying any loss twice.
Compare stack demand with logic, substrate and assembly constraints before predicting delivered accelerators.

Science and assumptions

Two alternative demand models

\begin{aligned}\text{Geometric:}\quad W&\approx\frac{B_{\mathrm{target}}a_{\mathrm{bit}}}{A_{\mathrm{usable}}Y}\\\text{Empirical:}\quad W&=B_{\mathrm{target}}I_{\mathrm{emp}}\end{aligned}

The geometric model divides target bits times gross area per bit by usable wafer area and an explicitly separate survival fraction. The alternative empirical model multiplies target output bits by measured wafers per output bit.

These are alternative models. An empirical output-bit intensity may already include yield; do not add the same loss again.

Symbols and units

W: Wafer requirement under the selected alternative model
wafers
B_target: Target delivered output bits
bits
a_bit: Gross silicon area per bit under the geometric convention
mm²/bit
A_usable: Effective usable area per wafer
mm²/wafer
Y: Explicit nonoverlapping survival fraction excluded from a_bit
fraction
I_emp: Empirical wafer requirement per delivered output bit
wafers/output bit

$B_{\mathrm{target}}$ is target output bits, $a_{\mathrm{bit}}$ gross silicon area/bit, $A_{\mathrm{usable}}$ effective wafer area, Y the explicitly modeled nonoverlapping survival, and $I_{\mathrm{emp}}$ wafers/output bit.

Assumptions Account separately for packing, redundancy and logic-process base dies; do not combine these alternative formulas blindly.

Engineering example

Synthetic empirical case: two wafer-equivalents per unit of tested-stack bits and ten units of demand require twenty wafer-equivalents. Applying a stack yield already embedded in that intensity would overstate demand.

Distinctions and tradeoffs

Wafer-supply intensity is neither package footprint nor a universal multiplier across HBM generations.

More capacity or bandwidth can consume additional wafer and assembly resources; architecture and memory-traffic improvements can reduce demand.

HBM can limit deployable accelerators even when logic wafers are available, connecting model memory requirements to manufacturing allocation.

Specifications and interfaces

Geometric method: Area per gross bit plus explicit yields
Separate memory and logic-process base-die wafer requirements.
Empirical method: Wafer supply per delivered bit
State generation, node, output boundary and included losses.
Different quantities: Footprint, summed silicon area and delivered capacity
Stack height and package area alone cannot measure wafer demand.

Interfaces

Memory requirement and scoped model Input: Target bits, product configuration and nonoverlapping yield/intensity assumptions
Normalized manufacturing demand Output: Wafer-equivalent or good-stack requirements over a stated period

Check your understanding

When is applying an extra stack-yield penalty to an empirical wafer/output-bit ratio justified?

Never for any geometric model either
Only when that ratio demonstrably excludes the stack loss
Always, because more factors improve accuracy

Answer and reasoning

Only when that ratio demonstrably excludes the stack loss

The starting metric’s denominator and included losses determine which additional factors are legitimate.

Evidence and scope

Micron’s March 20, 2024 prepared remarks, page 6, report approximately 3× wafer supply for HBM3E versus DDR5 at equal bits and the same technology node. This historical supplier statement does not establish a universal generation-independent factor, package footprint or current customer allocation. The lab keeps synthetic inputs and does not add losses already embedded in an empirical ratio.

Compare a synthetic geometric area ratio with an output-normalized supply ratio without applying the same yield loss twice.