Reference library
More actions
The guides, reviews, schematics, papers, talks, research notes and benchmark sites that members point newcomers and researchers to, with what each one covers, how current it is and who pointed to it. Where an outside source and a wiki page disagree, the wiki page carries the newer measurements and its sources. Sources that are AI-generated or unvetted are listed under Unvetted or AI-generated sources so they can be recognised.
Papers and talks about how the unlock was found are listed as pointers only. This wiki describes what the published tools do and how to check they worked; it does not carry keys, decrypted firmware or exploit details. Material that would reproduce one person's long write-up waits for the author.
Start here
The cmpunlocker GitHub is the first-party starting point
On 2026-09-07 Waka324 told newcomers that everything relevant is in the Discord and that the Gen3 unlock was not available. After the LTT video brought in newcomers on 2026-09-12, regulars told them to read the cmpunlocker GitHub (amoghmunikote/cmpunlocker) first: it lists the common issues and what can and cannot be done yet. insomniac suggested having an AI summarize it; Far said to read the first-party source directly because "It's already organized." DaLZy also pointed people to NVIDIA's GA100 whitepaper. On this wiki: Driver unlock and Cmpunlocker forks.
Reported
Source: @cmdrwaka324, 2026-09-07
Source: @ins_mniac, 2026-09-12
Source: @farawaythrower, 2026-09-12
Source: @dalzyyy, 2026-09-12
The 170th Street GitBook is the community guide and FAQ
On 2026-08-20 a new owner said there were "so many forks going around lately" and asked for one path covering unlock, firmware and running models. ST pointed him to the amoghmunikote/170th-Street repository, the source of the 170th Street GitBook (170th-street.gitbook.io/hx), whose embed describes it as the most comprehensive community guide. On 2026-09-21, when Bjerke asked for a FAQ, ST pointed to the same repository; Lycorascence noted that consolidated information may change as new breakthroughs land.
Reported 8 GB and 10 GB
Source: @stcord, 2026-08-20
Source: @unmotivatedgene, 2026-08-20
Source: @stcord, 2026-09-21
Source: @bjerke_, 2026-09-21
Source: @lycorascence, 2026-09-21
What the 170th Street GitBook covers
Eight pages by Amogh, the author of cmpunlocker: What is the CMP 170HX, History and Timeline, Full Specifications, Teardown Guide, Leaked NVIDIA A100 Schematic, Current Unlock, PCIe Capacitor Mod and Watercooling Installation. The specifications page identifies the variants (8 GB card: GA100-105F-A1, PCI device 0x20C2, 4096-bit HBM2e; 10 GB card: GA100-105A-A1, PCI device 0x2082, 5120-bit) and lists the stock restrictions measured before the unlock (FP32 0.41 TFLOPS throttled against 12.99 unlocked, PCIe Gen1 x4, Resizable BAR limited to 64 MiB, ECC disabled). The Current Unlock page (July 2026) describes cmpunlocker as a driver-side path applied each time the patched modules load, restoring 64 GB on the 8 GB card and 40 GB on the 10 GB card, unthrottled compute and PCIe Gen2, and marks it highly experimental. The repository also bundles the A100 reference schematic PDFs, the envytools docs and raw board photos (the clone is about 214 MB). It has no license file, and its images are hotlinked from the niconiconi review. On this wiki: Which card do I have, Variant comparison, PCIe x16 capacitor mod, Water cooling and Disassembly and cleaning.
Reported 8 GB and 10 GB
Source: @mrmannyatx, 2026-08-02
Source: @mrwojaketh, 2026-08-03
Source: @fipzy, 2026-09-11
Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18
Source: @darkmaniac7, 2026-10-30 (lab notes)
170th Street: inconsistent SM counts and a July 2026 snapshot
The intro page gives the 8 GB card 3584 cores, 56 SMs and 224 tensor cores, and the 10 GB card 1865 GB/s, while the specifications page gives 70 SMs and 4480 cores for both cards and 1.49 and 1.56 TB/s. The last commit (2026-07-30, "Update bad SM count") shows this was being corrected. The guide predates the RIR fuse burns, the Gen3 work and the cmpunlocker forks, so check anything about the unlock against Driver unlock, RIR fuse burn and Cmpunlocker forks.
Reported 8 GB and 10 GB
Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18
Source: @darkmaniac7, 2026-10-30 (lab notes)
170th Street: the llama.cpp page works, the LLM-bench section is outdated
On 2026-07-19 Wolsani pointed newcomers to the GitBook's llama.cpp inference page for a quick LLM test, now with a larger model since memory is unlocked, and said the LLM-bench section is outdated but its scripts still work. Amogh, who maintains the guide, said it had not been updated in some time, and coolgod asked whether the community could help edit it. The same thread linked the GitBook's NVLink population research page. Engine notes on this wiki: llama.cpp on CMP cards and NVLink.
Reported 8 GB and 10 GB
Source: @_j3ns, 2026-07-19
Source: @wolsani, 2026-07-19
Source: @coolgod_72592, 2026-07-19
170th Street: early material was removed as inaccurate (2026-07-28)
Phoenix17 noted that removed content is still in the repository's commit history. Amogh said he removed it because it was "highly inaccurate", dating from the early, assumption-based phase of the project, and that he was putting correct information back slowly. Treat claims found in old commits as superseded.
Reported
Source: @_j3ns, 2026-07-28
Source: @phoenixone7, 2026-07-28
Reviews, teardowns and hardware documents
niconiconi: "All GB/s without FLOPS" review and teardown
niconiconi's review "All GB/s without FLOPS - Nvidia CMP 170HX Review, Performance Lockdown" on niconiconi.neocities.org is the standard teardown and measurement reference from before the unlock. MrManny shared it on 2026-08-02 for its teardown photos, and DaLZy later linked it as a reference for diagnosing a dead card. Its repair appendix circulates as a separate PDF (next entry but one). On this wiki: Recovery and Disassembly and cleaning.
Reported 8 GB and 10 GB
Source: @mrmannyatx, 2026-08-02
Source: @dalzyyy, 2026-08-07
A 170HX PCB photo with three added wires, and the LTT teardown
quark2world linked a CMP 170HX PCB photo that the 170th Street GitBook pulls from the niconiconi review. The photographed card has 3 wires routed over from the power connector, and he thinks the card in Linus Tech Tips' video had the same, possibly a monitoring mod. When neko doubted the GPU existed, quark2world pointed to the LTT video, which includes a teardown. Both messages are from 2026-07-23, before LTT's 2026-09-12 video, so they likely mean LTT's earlier coverage of the card (2021-11-25); not confirmed. Coverage list: Press and coverage.
Reported
Source: @quark2world, 2026-07-23
Source: @nekopavel, 2026-07-23
No boardview exists; the A100 reference schematic and the hardware-repair thread
On 2026-08-02 DaLZy asked whether a boardview exists for the A100 40G or the CMP 170HX; Amogh: "Unfortunately not, I wish we did." SealClubber pointed to the only available document, the A100 GA100 reference schematic (NVIDIA-A100-GA100-883-P1001-B02-Rev-A.pdf), kept in the 170th-Street repository. On 2026-09-11 fipzy started a hardware-repair thread with the same schematic, which he says also suits the CMP 170HX, plus a repair appendix (Appendix_3_Repair.pdf) copied from the niconiconi review and power-connector images. On this wiki: Recovery and PCIe x16 capacitor mod.
Reported 8 GB and 10 GB
Source: @dalzyyy, 2026-08-02
Source: @_j3ns, 2026-08-02
Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @sealclubber, 2026-08-02
Source: @fipzy, 2026-09-11
A100 die underside photos; SXM-to-PCIe reballing exists as a service
quark2world posted two photos of the underside of an A100 chip on 2026-07-23. They come from an X post by regsky_, who offers reballing from SXM to PCIe boards as a service and is known in the US scene for 4080 and 4090 upgrades.
Reported
Source: @quark2world, 2026-07-23
Source: @roger_dammit, 2026-07-23
DGX Station A100 cooling reference
On 2026-08-10 quark2world shared a Reddit thread on overheating DGX Station A100s and NVIDIA's DGX Station A100 architecture white paper as a watercooling reference for A100-class parts. bayley_w described the design as a low-pressure evaporative cooling loop with a Redstone SXM4 board hanging at the back. On this wiki: Water cooling and Thermal limits.
Reported
Source: @quark2world, 2026-08-10
Source: @bayley_w, 2026-08-10
Cloud A100 PCIe node topology screenshots
JJ tested cloud A100s on 2026-09-08 while waiting for his cards and found a node with 4 PCIe cards "in 3x16"; he posted topology screenshots, likely of NVLink-bridged pairs, and 大乐 posted another. A possible A100 PCIe and NVLink topology reference, but the details are only in the images. On this wiki: NVLink and P2P.
Reported
Source: @jj48., 2026-09-08
Source: @dale063629, 2026-09-08
Unlock research: background reading
Jon Pry: "A Canary in the Crypto Mine"
On 2026-08-23 meatsus recommended Jon Pry's paper as the starting point for understanding NVIDIA's mining-card restrictions and how the CMP unlock was approached. He noted it needs at least GitHub use and compiling C from source, and suggested asking an AI assistant to explain it for beginners. On 2026-08-24 xanthe_ pointed a new reverse-engineering volunteer to the same paper, "A Canary in the Crypto Mine: Defeating Stack Protection in a GPU Secure Coprocessor". On 2026-09-30 Lycorascence linked it on ResearchGate as the document people had quoted. History of the unlock: Unlock history.
Reported
Source: @meatsus, 2026-08-23
Source: @xanthefin, 2026-08-24
Source: @lycorascence, 2026-09-30
A reading order for newcomers to the unlock research (2026-09-30)
mrpandagod's suggested order:
- Watch the seminar talk by Petri.
- Read the commit history in all cmpunlocker branches, including deleted comments and files, to learn how it works and which registers are which.
- For memory, read the initial paper, then the Chinese SAM microscopy paper.
- Do non-destructive fuse-status testing and cross-reference it with legal, non-proprietary information.
He had already pointed newcomers to a "seminar talk" for background on 2026-09-22. Read-only fuse tools and maps: Gpcprobe and Fuses and floorsweep maps; the tool that writes fuses is on RIR fuse burn.
Reported 8 GB and 10 GB
Source: @djpandaman, 2026-09-22
Source: @djpandaman, 2026-09-30
Source: @lycorascence, 2026-09-30
Petri Krohn's MRMCD2026 talk
On 2026-09-26 PJ pointed to the MRMCD2026 conference talk "FACEB13D exploit: Liberating the A100 beast inside Nvid..." (YouTube) as the best direction to explore. On 2026-10-24 Petri Krohn linked the same video as the recording of his presentation, and lab notes list it as his unlock-history talk. Whether the "seminar talk" in mrpandagod's reading order is this talk is not confirmed.
Pointer: the talk is not summarized on this wiki; a summary waits for Petri Krohn. See Discord.
Reported
Source: @zingy90210, 2026-09-26
Source: @petrikrohn_26497, 2026-10-24
Source: @darkmaniac7, 2026-10-24 (lab notes)
CMP_170HX_unlock_September_2026_2.pdf
On 2026-09-26 PJ posted an 8.4 MB PDF named CMP_170HX_unlock_September_2026_2.pdf. On 2026-09-28 he posted it again in #suppliers-chat with three slide screenshots, saying Petri Krohn is one of the people who made the unlock and worked alongside Jon Pry. On 2026-09-29 he posted it once more; Humvee55 replied that Petri is in the server and "we all work together", which suggests Petri wrote it. Its contents were never summarized in the chat. It is likely a primary source for the unlock history.
Pointer: the PDF is not hosted, quoted or summarized here until Petri Krohn confirms he wrote it and agrees. See Discord.
Reported 8 GB and 10 GB
Source: @zingy90210, 2026-09-26
Source: @zingy90210, 2026-09-28
Source: @zingy90210, 2026-09-29
Source: @scallion_9883, 2026-09-29
Research notes, gists and blogs
Research material collected in lab notes on 2026-10-24 and in the older Discord history. These are notes, not procedures: check a claim against the wiki page it belongs to before acting on it.
Consensus-Protocol/cmp170hx: an AI-generated research wiki (unverified)
Thanos built the Consensus-Protocol/cmp170hx GitHub wiki entirely with AI from roughly 30,000 Discord messages across two servers plus papers, gists and code references, and said on 2026-08-01 that it is not verified. He later called it a "450 agent workflow" with no names attached. CUTE found it AI-written but correct at first glance, with interesting NVLink material, and Danske suggested a server bot for commands, guides and scripts. Humvee55 praised its history write-up; clee found its Frontier-80GB page inconsistent about the 80 branch; Studebaker suggested it as a knowledge source for the server bot, noting it likely duplicates information. On 2026-08-11 Humvee55 pointed to it as the wiki that "explains all the history etc on how the original compute unlock worked", Matty Ice linked it, and Josephur read its compute-throttle page to check whether it applies to the CMP 100-210. In October members linked its PCIe Gen3 and Gen4 frontier page and its VBIOS version inventory; lab notes give its last commit as 2026-08-22.
Pointer: its pages are not copied here; reuse waits for Thanos. Treat each claim as unverified until a wiki page confirms it with a source.
Reported 8 GB and 10 GB
Source: @thanos8912, 2026-08-01
Source: @cute_sc2, 2026-08-01
Source: @danske1, 2026-08-01
Source: @_studebaker_, 2026-08-06
Source: @thanos8912, 2026-08-06
Source: @scallion_9883, 2026-08-06
Source: @c133, 2026-08-06
Source: @scallion_9883, 2026-08-11
Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @icematty, 2026-08-11
Source: @josephur, 2026-08-11
Source: @rrickk420, 2026-10-09
Source: @zagitta, 2026-10-12
Source: @darkmaniac7, 2026-10-24 (lab notes)
club-170hx: notebooks, QC scripts and results
PixelML's club-170hx repository holds notebooks (among them a 4-card BAR1 P2P run behind a PLX switch, 2026-10-01), the QC scripts and results that several wiki pages cite. hum1d planned to validate three incoming cards with its steps. On this wiki: P2P, PCIe switch boards, Acceptance checklist and Testing and burn-in.
Reported
Source: @hum1d, 2026-10-07
Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @djpandaman, 2026-10-22
Source: @darkmaniac7, 2026-10-24 (lab notes)
ga100-fuse-re: GA100 fuse reverse engineering
ahnguyen17's ga100-fuse-re repository documents the GA100 (CMP 170HX) fuse architecture. eXile linked it on 2026-10-07 in a question about RIR burns on the 10 GB card, and lab notes use its findings in their reading of the RIR tool. On this wiki: Fuses and floorsweep maps and RIR fuse burn.
Reported
Source: repo github.com/ahnguyen17/ga100-fuse-re by ahnguyen17 on GitHub, posted by @exile_1ck, 2026-10-07
Source: @darkmaniac7, 2026-10-24 (lab notes)
GA100 VBIOS comparison tables (the nvlight gist and Humvee55's fuses gist)
The nvlight gist is a GA100 VBIOS comparison table (CMP 170HX against A100 PCIe and Drive A100) with dump and parse scripts; lab notes treat it as the reference for any CMP VBIOS work. An earlier copy of the comparison table sat on 2026-07-22 in JRex286's fork of Humvee55's fuses gist, next to probe.sh, which reads fuse values and VBIOS data from a running card (nvflash alone is enough for a VBIOS dump). Humvee55 stressed that the compute override "lives in fuses, not vbios", so fuse dumps are needed to study it. On this wiki: VBIOS and firmware and Fuses and floorsweep maps.
Reported 8 GB and 10 GB
Source: @scallion_9883, 2026-07-22
Source: @serplay228, 2026-07-22
Source: @darkmaniac7, 2026-10-24 (lab notes)
blog.kkk.rs: HBM and memory-unlock analysis of the 10 GB card (in Chinese)
blog.kkk.rs/archives/71, "170Hx 10G解锁80G探索" ("10G unlock to 80G, an exploration", title translated from Chinese), analyses the HBM stacks and the memory unlock on the 10 GB card. On 2026-08-11 quark2world posted an LLM summary of it: the author used an old NVIDIA MODS diagnostic script, hbm.jse, which works only with the 455.263.1 driver branch, to read the HBM device IDs; the Samsung stacks reported "XA2_8HI", which the author matched in driver source to Samsung X-die rather than the B-die used on V100. That reading is secondhand; check it against the post. Where the 10 GB card stands on memory today: Memory unlock on the 10 GB card and The 10 GB card: provenance and the 80 GB question.
Reported 10 GB
Source: @meromgreen, 2026-08-11
Source: @quark2world, 2026-08-11
Source: @djpandaman, 2026-10-12
Source: @darkmaniac7, 2026-10-24 (lab notes)
matstatman's 0xBAD reference gist
Per-HBM-stack and TPC temperatures are hidden on these cards: reads return 0xBADF codes (0xBADF1005, 0xBADF1401, 0xBADF1002), which giaki3003's agent decoded as FECS PRI timeout and stuck-ack codes. matstatman published a 0xBAD reference gist for reading such values. CUTE says the sensors are hidden or firmware-reserved, not fused; matstatman thinks they are fuse-controlled. Miséricorde noted that monitoring shows only the hottest HBM stack; Hbmmon later read per-stack HBM temperatures. On this wiki: Monitoring.
Reported
Source: @matstatman, 2026-09-19
Source: @wakeupmbrc1, 2026-09-19
Source: @cute_sc2, 2026-09-20
Source: @giaki3003, 2026-09-20
Videos
RH-3D's tuning videos for the 8 GB card and the x16 mod video
RH-3D keeps a YouTube series for the 8 GB card: "#011 Ultimate Tuning" (YouTube), shared as the guide to undervolting with cmpunlocker; "#012" (YouTube), a follow-up after the ECC merge; and "Cmpunlocker v0.5 for Nvidia CMP170HX 64GB VRAM +4SMs +ECC SRAM" (YouTube). On 2026-09-23 mrpandagod recommended an earlier rh3d video on the x16 capacitor mod (YouTube) as the hands-on companion to the 170th Street guide, after rickflare found the guide does not explain the soldering technique, and advised against trying the mod without the skills. No jig for the mod was mentioned. On this wiki: Undervolting and power limits, Undervolting guide: where it came from and PCIe x16 capacitor mod.
Pointer: the videos are not transcribed here; a written guide from them waits for RH-3D. See Discord.
Reported 8 GB
Source: @djpandaman, 2026-09-23
Source: @.rickflare, 2026-09-23
Source: @rh3dai, 2026-10-04
Benchmarks and recipe collections
llm-inference-bench, tool-eval-bench and localairecipes.com
Two benchmark harnesses come up for these cards: llm-inference-bench, a serving benchmark over a concurrency and context matrix that Mountains linked on 2026-10-12, and SeraphimSerapis' tool-eval-bench, which Nestaa51 used beside it for a GLM 5.3 Flash against Qwen3.8 Flash Next comparison on two cards (2026-10-14). Mountains' localairecipes.com went public on 2026-10-06 with CMP 170HX leaderboards for one to four cards; much of its data is self-reported, so verify before relying on a figure. Results on this wiki: Benchmarks and All recipes.
Reported
Source: @mountains1605, 2026-10-06
Source: repo github.com/local-inference-lab/llm-inference-bench by local-inference-lab on GitHub, posted by @mountains1605, 2026-10-12
Source: repo github.com/SeraphimSerapis/tool-eval-bench by SeraphimSerapis on GitHub, posted by @nestaa51, 2026-10-14
Source: @darkmaniac7, 2026-10-24 (lab notes)
Unvetted or AI-generated sources
Listed so they can be recognised. Do not act on them without a wiki page or a second person confirming. The Consensus-Protocol wiki above belongs here too: AI-generated and unverified by its author's own account.
The infinite.im SM table: AI-written, with known errors
Xeax's table on infinite.im lists the active SMs of NVIDIA RTX and CMP cards, so that inactive ones could in theory be re-enabled if they are not defective. emogop questioned its row saying a 90HX can reach all 84 SMs while a 3080 cannot; Xeax said that row came from his AI and asked for corrections. Lab notes flag the table as AI-written with known errors. To see what a card actually has, start with Gpcprobe (matstatman's advice in the same thread).
Reported
Source: @predictorandrzej, 2026-10-18
Source: @xeax, 2026-10-18
Source: @matstatman, 2026-10-18
Source: @emogop, 2026-10-18
Source: @darkmaniac7, 2026-10-24 (lab notes)
morluto/rea: an agent reverse-engineering tool, hyped and unvetted
Nestaa51 posted morluto/rea ("reverse engineer anything with agents") on 2026-10-09. Mountains noted it lacks NVIDIA Falcon support, several people found it suspicious because it was being pushed on X and YouTube feeds, and lab notes call it hyped and unvetted. Nobody reported using it successfully on these cards.
Theory
Source: @nestaa51, 2026-10-09
Source: @mountains1605, 2026-10-09
Source: @darkmaniac7, 2026-10-24 (lab notes)
The mintoriakamoto CMP 100-210 repository: its author was banned here
The mintoriakamoto CMP100-210 tensor-only repository came from a developer whose claims are flagged as unreliable. In the order posted:
- JMA1976 ran its tensor-only mode (boot-time, reapplied after each boot) on his 24 CMP 100-210 cards without enabling Gen3 or flashing the VBIOS: a large performance gain at low power and 5 days without issues. Source: @jma1753, 2026-10-09
- 68x warned that the same developer claimed PCIe 5.0 on the CMP 170HX in cmpunlocker issue #9 and called the work AI slop; Amogh confirmed he banned that person. Source: @68x0, 2026-10-09
Source: @_j3ns, 2026-10-09
Community CMP 100-210 unlocks with a track record: Other CMP cards and other NVIDIA GPUs.
Contested
History and side reading
Why the CMP 170HX exists: origin theories
In the order posted on 2026-09-18:
- Takka: a lower-binned A100 meant as an export-legal option for China, never exported (possibly because US rules tightened), then crippled into a mining card. He estimated that at most 100k to 200k 170HX were made against millions of A100s, assuming serial numbers are exhaustive. Source: @spectakkular, 2026-09-18
- xtj7: the 170HX launched in 2021, while China AI export restrictions came in 2022. NVIDIA crippled A100 silicon where it would not hurt mining, to avoid cannibalising A100 sales during the mining boom, before ChatGPT changed demand. Source: @xtj7, 2026-09-18
- Miséricorde: the A100 market was saturated. Source: @wakeupmbrc1, 2026-09-18
- cibernox: the cards became e-waste after Ethereum moved to proof-of-stake, which fits the low days-of-use counters people see in their BIOSes. Source: @cibernox, 2026-09-18
Dated events: Unlock history and Press and coverage.
Contested
Intel Battlemage needs a BAR1 resize; the 4028GR BAR1 P2P kernel trick
bayley_w (2026-08-13): Intel Battlemage (B-series) drivers no longer start, even for compute, unless they can resize BAR1 during initialization; without BIOS support, OpenCL and SYCL contexts refuse to initialize. He says the same kernel-patch trick used for BAR1 P2P on the Supermicro 4028GR server would fix it, though it is a lot of effort for a $100 GPU. On this wiki: Supermicro SYS-4028GR-TRT2 and Resizable BAR and BAR1 64 GB.
Reported
Source: @bayley_w, 2026-08-13
Source: @cmdrwaka324, 2026-08-14
The BC250 side project and the LTT video
On 2026-09-01 Studebaker linked the cachenetics/project-ariel BC250 patches for CachyOS kernel 7.0.9 and a BC250 manual, noting it still needed updating for the 8c/16t unlock. Rickcal said BC250 prices doubled after a video (one with a 1 TB SSD for $375). On 2026-09-05 Petri Krohn noted that Linus Tech Tips had published a BC250 unlock video; LT said it had been an open secret for 6 months to a year. Other hardware at these prices: Alternatives at these prices.
Reported
Source: @_studebaker_, 2026-09-01
Source: @rrickk420, 2026-09-01
Source: @petrikrohn_26497, 2026-09-05
Source: @licensedterrapin, 2026-09-05