Test whether the "defective" 8 GB HBM stacks are usable at lower voltage and clocks under water

Status: open
Priority: 0
Reference: discord:1558184726261399553
Public link
Back to the board

Swimlane: Default swimlane
Column: Wanted
Position: 28

Assignee: not assigned
Creator: cmp.wiki

Started:
Created: 10/30/2026 09:12
Modified: 10/30/2026 09:12
Moved: 10/30/2026 09:12

8 GB

Description

package-height binning theory

Why: 80/96 GB on 8 GB rests on it

Reported by / on it: @n2dcrypt, @globalapathy

Source: https://discord.com/channels/1518634379604136178/1518695328008638516/1558184726261399553

Wiki page: 8 GB card

Cards: 8GB

Card T-029, seeded from the cmpunlocker Discord with credit (reference discord:1558184726261399553).

Sub-Tasks

Internal links

Comments

cmp.wiki Created at: 11/06/2026 06:18 Updated at: 11/06/2026 06:18 Visibility: Standard users

Board audit 2026-11-06: partly answered, card stays open.

The lab's read-only floorsweep read of seven 8 GB cards found 2 FBPs per card fused defective and 2 more gated but not dead, with no volatile override (CTRL_OPT_FBP = 0). That settles which partitions the fuses mark defective, and the floorsweep page reads them as dead silicon, but it is a fuse read, not a test of the memory. Nobody can enable a marked-defective partition on the 8 GB card in software, so this test cannot be run yet, and the wiki keeps "The defective HBM stacks are good memory" Contested.

https://cmp.wiki/wiki/Floorsweep_read_of_the_8_GB_card:_gated_versus_defective_partitions
https://cmp.wiki/wiki/Rumours_and_claims_to_verify (section "The defective HBM stacks are good memory")