HBF does not replace HBM. It adds one tier to the memory hierarchy. Suppose that in one decode step a fraction f of the bytes read comes from HBF and the remaining 1−f from HBM. If the two are read one after the other, the time taken is the sum of the two shares.
The slider and the horizontal axis of the chart use a square-root scale to spread out small values of f.
Slider range = the span of the three grades in the public specification released at OCP in 2026-08 (about 0.4–3.0 TB/s). The value for the middle grade could not be confirmed from public sources.
HBM4 stack bandwidth BHBM
HBM4 stack capacity CHBM
24–36 GB is the Samsung volume-production 12-high stack; 64 GB is the standard maximum (32Gb dies, 16-high). The HBF stack capacity CHBF is fixed at the public spec maximum of 512 GB.
Effective bandwidth Beff
Relative to HBM-only reads
Total capacity C
HBF share of step time
Check against the text ()
Worked example: f = 0.1, HBM4 3.3 TB/s and 36 GB (Samsung announced maximum), HBF 0.4 TB/s (assumed value) and 512 GB. The "Text" column holds values from the worked example in Session 3, Section 6 ("Mixing the two tiers"), from "Status and specifications", and from the Section 4 table "Specifications by generation" and the list below it.
Item
Computed
Text
Data that can keep f small
Once the share read from the slow tier grows even a little, that tier sets the overall speed. The place for HBF is therefore data that can keep f small: data that is mostly held and only occasionally read, such as the MoE expert weights not used in the current step. Which data qualifies is sorted out in Session 3, Exercise 4 ("Judging HBF fit") along four axes: read/write ratio, reuse distance, access granularity and latency sensitivity. The one-line rule is: "Is it read-only and large, and can you predict in advance when it will be needed?"
What this calculation leaves out
The model assumes the two tiers are read in sequence. Overlapping reads from both at once, and moving needed data into HBM ahead of time (prefetch), are not included.
The long latency of NAND is also folded into a single bandwidth number.
All bandwidth values are announced or assumed figures for a single stack.
As of October 2026 the first HBF standard specification has been published, but no volume-production product exists; the performance figures are spec ranges and targets.