Memory unlock on the 10 GB card: Difference between revisions

Revision as of 08:49, 30 October 2026

Memory on the 10 GB card (40 GB, 48 GB on a lucky card)
Reported 10 GB
Driver	cmpunlocker-patched driver (610+)
Unlock build	cmpunlocker (40 GB) + RIR software override (6th stack)
Steps	Memory unlock on the 10 GB card
Source	link
Confirmations	None yet.

The 8 GB card unlocks to a flat 64 GB (see Driver unlock). The 10 GB card is different and the ceiling is a lottery. Warning Do not expect 80 GB.

40 GB (the baseline)

cmpunlocker sets the 10 GB card to 40 GB (40960 MiB): CFG1 0x009a0204 = 0x02669000, LMR 0x00100ce0 = 0x0000028A. That is 5 HBM stacks of 8 GB. This is Proven and reproduced by everyone running the unlock. Source: @programmerboi, 2026-10-25

48 GB (the 6th stack, if it is good)

48 GB comes from enabling a disabled sixth stack (48 GB = 6 x 8 GB). It needs the RIR fuse burn (software fuse remask) and the card must have a working 6th stack, which not all do; estimates put the chance around 20% or less. Source: @djpandaman, 2026-10-11 Source: @artistosu, 2026-10-18 On most sampled cards the last stack is marked defective, not merely disabled (bigfoot: all 6 of his; orrorin: both of his), and defective-marked stacks really are defective. A "pristine" card with the 6th stack only disabled (gpcprobe OPT_FBP_DISABLE 0x840, all DEFECTIVE masks 0) is the one that can reach 48 GB. Source: @bigfoot1144, 2026-10-19

HBM vendor

Reported The 10 GB card carries Samsung HBM2e; the 8 GB card carries SK hynix. Checked by Ivan on his own cards. Source: @darkmaniac7, 2026-10-30 (lab notes)

Earlier Discord posts on the question, kept for reference:

RH-3D: 10 GB = Samsung HBM2, 8 GB = Hynix HBM2e 16 GB stacks.

 Source: @rh3dai, 2026-10-17

atsasin: the 10 GB card never had HBM2e, only the 8 GB.

 Source: @atsasin2126, 2026-10-16

mrpandagod: the evidence says the stacks are binned 16 GB HBM2e.

 Source: @djpandaman, 2026-10-17

Mark Linton's read-only dumps of two 8 GB cards show the SK Hynix strap 0x004906D1.

 Source: @marklintonca, 2026-10-10

80 / 96 GB: not real on these dies

Contested Several people got a 10 GB card to report 80 GB, but the memory was unstable above 40 GB. Source: @tekkenz, 2026-10-12 Humvee55, a co-author of the paper that used a 10 GB card as the test vehicle, now believes the apparent extra capacity is aliasing in L2, not real memory. Source: @scallion_9883, 2026-10-16 A read-only fuse RE of a Samsung 10 GB card found the 80 GiB aperture (LMR 0x28B) folds at about 16 KB above the expected address (address folding confirmed die-side), i.e. the upper half is aliased, not addressable. Source: @ahnguyen17, 2026-10-03 Stacks are physically 16 GB parts that Samsung binned; unlocking the upper half is "basically impossible" with the information available, likely needing internal Samsung codes. Source: @djpandaman, 2026-10-11

A Chinese write-up "170Hx 10G 解锁80G探索" (10G unlock to 80G exploration) is at blog.kkk.rs/archives/71; no one reported a stable 80 GB card from it. A screenshot claiming 130104 MB was debunked as two 64 GB-unlocked 8 GB cards reported together (65024 MiB x 2). Source: @kio063218, 2026-10-15

How the state is reported

Community agreement as of 2026-10-23: a 10 GB card now equals "an A100 40G PCIe with more SM"; after the RIR tool went open source, except for the 40 to 80/96 GB VRAM-initialisation issue, most unlock problems are solved. NVLink is still not public. Source: @ob88594, 2026-10-23

Known issues

A dead region that returns 0 and stores nothing: AnalyticETH had an agent write a script that pins a 2 MB throwaway

 allocation over the bad address at startup so nothing important lands there, rather than discarding the card.
 Source: @analyticeth, 2026-10-08

Cards with HBM errors: Zombie's card threw about 5,000 single-bit and some multi-bit errors per 5-minute

 memtest_vulkan run, always at the same addresses (around 30 GB and 61 GB). He mapped the failing
 addresses to physical memory and used the driver's boot-time bad-block marking; only 4 blocks (64 KB each) were bad,
 costing 8 MB (63.06 to 63.05 GB), after which memtest_vulkan and gpu-burn passed with ECC
 off. The script must run every boot; the card refuses if there are too many errors (multiple GB). See
 Troubleshooting. Source: @zombielurker, 2026-10-05

Status

40 GB is Proven. 48 GB is Reported and card-dependent. 80/96 GB is Contested and most evidence says it is aliasing, not memory.

@@ Line 69: / Line 69: @@
 == Known issues ==
-* A dead region that returns 0 and stores nothing can be worked around by pinning a dummy allocation over it; a longer
+* A dead region that returns 0 and stores nothing: AnalyticETH had an agent write a script that pins a 2 MB throwaway
-   write-up by @analyticeth exists; see Discord.
+  allocation over the bad address at startup so nothing important lands there, rather than discarding the card.
-* Cards with HBM errors: the driver can mark bad physical blocks at boot (4 x 64 KB blocks cost 8 MB). A longer write-up
+   {{Source|analyticeth|2026-10-08|https://discord.com/channels/1518634379604136178/1518692084729511967/1557566171195383850}}
-   by @zombielurker exists; see [[Troubleshooting]] and Discord.
+* Cards with HBM errors: Zombie's card threw about 5,000 single-bit and some multi-bit errors per 5-minute
+  <code>memtest_vulkan</code> run, always at the same addresses (around 30 GB and 61 GB). He mapped the failing
+  addresses to physical memory and used the driver's boot-time bad-block marking; only 4 blocks (64 KB each) were bad,
+  costing 8 MB (63.06 to 63.05 GB), after which <code>memtest_vulkan</code> and <code>gpu-burn</code> passed with ECC
+   off. The script must run every boot; the card refuses if there are too many errors (multiple GB). See
+  [[Troubleshooting]]. {{Source|zombielurker|2026-10-05|https://discord.com/channels/1518634379604136178/1518692084729511967/1556530076105838723}}
 == Status ==