Full SM unlock

Full SM compute cores (up to 128 on the 10 GB card)
Proven 10 GB
Driver	cmpunlocker-patched driver (610+)
Unlock build	170hx-rir-tool + software IFF
Steps	Full SM unlock
Source	Discord message
Confirmations	Reproduced by @ghostinthedatacenter_87281 (10 GB), 2026-10-19 · details

The stock card exposes 70 of 128 SMs. There are three tiers:

74 SM on both cards: cmpunlocker's +4 SM step (70 + 4). This is the ceiling for the 8 GB card.

 Source: @lia_oc, 2026-10-22

74 to 80 SM on the 10 GB card: software only, no fuse override needed.

 Source: @djpandaman, 2026-10-10

80 to 128 SM on the 10 GB card: needs EN_SW_OVERRIDE, which needs the RIR fuse burn.

Why the 8 GB card cannot go past 74

The extra TPCs beyond 74 are controlled by OPT_TPC_DISABLE, which is burned to 1 on the 8 GB card, and there is no free RIR to override it. Source: @djpandaman, 2026-10-10 On 2026-10-26 riko said there is "no pub unlock for add more SM for any cmp170 8-10g except 70+4 sm". Source: @riko6253, 2026-10-26 When GhostInTheDatacenter showed a 10 GB card that started from 170hx-rir-tool, riko asked "but there is no sm unlock isn't it?"; 68x0 answered "You can unlock more SMs with the RIR burn", and GhostInTheDatacenter confirmed. So the only public path past 74 SMs is the RIR burn on the 10 GB card. Source: @riko6253, 2026-10-26
Source: @68x0, 2026-10-26
Source: @ghostinthedatacenter_87281, 2026-10-26

10 GB: how far it goes

mrpandagod took a single 10 GB card from 74 to 80 to 94 to 108 to 122 SM, FP16 roughly 12-14 to 22 TFLOPS. Source: @djpandaman, 2026-10-09 lxwx got the last cores by overriding OPT_TPC_DISABLE ("just override OPT_TPC_DISABLE"), then reached 128 SM at full clock. Source: @lxtend, 2026-10-12 Source: @lxtend, 2026-10-17 On 2026-10-17 Feralzi reached 126 SM and Waypoint enabled every SM not marked defective (see 10 GB card). Source: @feralzi, 2026-10-17
Source: @waypoint3642, 2026-10-17 GhostInTheDatacenter posted a screenshot of an unlocked 10 GB card on 2026-10-26, said it was unlocked in software, started from 170hx-rir-tool and was "really just giving info to my clanker until it got it right". Source: @ghostinthedatacenter_87281, 2026-10-26
Source: @ghostinthedatacenter_87281, 2026-10-26
Source: @ghostinthedatacenter_87281, 2026-10-26 orrorin's own report is lower: both of his 10 GB cards stop at 80 SMs (see the next section). Source: @_orrorin, 2026-10-25 clpeak on mrpandagod's card at 122 SMs measured 21.9 TFLOPS FP32 and 43.7 TFLOPS FP16 on the vector units; the full table is on Benchmarks#Compute and memory peaks. Source: @djpandaman, 2026-10-23

How many SMs a 10 GB card reaches

Contested 10 GB Cards have landed anywhere from 80 to 128 SMs, and people disagree on why some stop low. In the order posted:

Every SM that is not defective. xeax: a 10 GB card can enable all the SMs on the chip that are not defective,

 up to 128 if you are very lucky (he puts the 8 GB card at 74 and the A100 at 108).
 Source: @xeax, 2026-10-25

Two cards stuck at 80. orrorin could unlock only up to 80 SMs on both of his 10 GB cards. He stepped up 2 SMs

 at a time, and anything above 80 failed a real model load; asked by xeax, he had also tried the other locked GPC.
 Source: @_orrorin, 2026-10-25

 Source: @xeax, 2026-10-25

 Source: @_orrorin, 2026-10-25

Operator error or binning. OB88594: it should be an operational failure, since unlocking more than 100 SMs is

 very common. xeax: more likely unlucky, heavily binned dies.
 Source: @ob88594, 2026-10-25

 Source: @xeax, 2026-10-25

Per-card masks

Asked whether the latest cmpunlock does this, xeax said no. The workflow is per card:

Unlock the card with the RIR tool (170hx-rir-tool, see RIR fuse burn).
Test which of the locked parts of the chip actually work.
Build a custom mask for that card that masks off the defective parts.

Source: @tugoubxs, 2026-10-25
Source: @xeax, 2026-10-25 xeax's untested sweep script is meant to automate step 2; see Tools.

The silicon lottery

GPCs/FBPAs marked defective are genuinely defective; enable only the disabled ones, and mask off the defective TPCs within a defective GPC. Source: @zorg33, 2026-10-22 Colo A/B-tested SM pairs and got 126 working SMs on a card where only one SM pair in a "defective" GPC actually failed. Source: @colo2442, 2026-10-22

Known issues

Warning Power cap above 80 SM. The moment an SM on a previously disabled GPC is enabled, the

 self-reported power becomes unreliable (readings above 400 W on a 250 W VBIOS) and the memory clock is forced down to
 about 1100 MHz with a large FP32 drop. Source: @djpandaman, 2026-10-13
 mrpandagod says he later resolved the wattage issue (fix not posted); the threshold starts at 80, not 90, which
 matters so people's agents are not misled. Source: @djpandaman, 2026-10-22
 See Cooling and power for undervolt / shunt options and Troubleshooting.

Early SM-unlock attempts crashed the card until the method was refined (lxwx, with help from mrpandagod by DM).
The method is derived from a leaked internal driver archive (integdev_gpu_drv.rar, 18.87 GiB, findable

 by file name on torrent sites) plus an agent, and from Humvee55's method notes (message.txt): his RIR
 plus software-IFF route, meant to be fed to an LLM to fill the gaps (see RIR fuse burn). lxwx's own contribution
 was finding how to run all 128 SM at full clock. Source: @lxtend, 2026-10-17

 Source: @scallion_9883, 2026-10-16

Status

128 SM reached and reproduced by several 10 GB owners, so Proven on the 10 GB card. The 8 GB card past 74 is Contested / no public path. See the confirmation below. How many SMs a given 10 GB card reaches is Contested (80 on some cards, see above).