Troubleshooting

Start here when a CMP 170HX does not do what it should. Find what you see in the first column, check the usual cause, and follow the link for the full story and the sources. Every fact on the linked pages is credited to the person who reported it; badges follow the Status taxonomy.

What you see

What you see	What it usually is	What to do	Details
The card is not in `lspci` at all	Slot, riser, power lead, or a damaged card (shipping damage, a dead regulator)	Full power-off, then change one thing per shutdown; try another slot and another system	Not detected
In `lspci` but missing from `nvidia-smi`	The card overheated with no airflow, the driver is 615.78.08, or GSP failed to boot	Put a fan on the card before powering on; use driver 615.71.09; read `dmesg`	Not in nvidia-smi
nvidia-smi shows 8192 or 10240 MiB after installing the unlock	No cold boot yet, or the stock driver module loaded instead of the patched one	Cold boot; check `dmesg \| grep SEC2_DEBUG` and `modinfo -n nvidia`	Still stock
Installer refuses, or the module build fails	Secure Boot on, missing kernel headers, a driver version not on the list, a kernel upgrade	Secure Boot off, headers for the running kernel, re-run after every kernel upgrade	Before you install
Link at x4 or x8 instead of x16	Stock is x4; after the capacitor mod, a cold or missing joint, a card not fully seated, or a slot that shares lanes	Read `lspci -vv` LnkSta, check each cap with a multimeter, reflow suspect caps, reseat the card	Link width
Link stuck at Gen1, Gen2 or below Gen3	Gen1 is stock and Gen2 comes with the unlock; Gen3 is fused off on 8 GB cards; one 10 GB card had a damaged PCIe controller	Check what your variant can reach before chasing it	Link speed, Unlocks
BAR1 stuck at 64 MB, "no space" for BAR1	The BIOS 64-bit MMIO window is too small (several ASUS boards, MS-01), or PLX switch firmware without hotplug	Above 4G Decoding, Resizable BAR, a large MMIO high window; mailbox P2P works at 64 MB	BAR1
An `Xid` line in `dmesg`	Depends on the number	Look it up	Xid errors
"GPU requires reset", `[N/A]` or NaN temperatures in nvidia-smi	A firmware crash (Xid 119, 154 or 62)	Power off for at least 30 seconds; a warm reboot is not enough	Recovery
GSP init fails on every boot	A failing card, a broken slot, or an eGPU dock in the path	Test in a native slot; see the open case	Xid errors (Xid 1), Recovery
Cards drop off the bus, or the whole rig loses power, on large prompts	Power: transients far above the cap, too many cards per PSU, shared cables	Lower the cap to 180 W, re-apply it after every driver load, give each card its own cables	Power, Power and cabling warnings
A 10 GB card with extra SMs power-throttles, or its memory clock drops to about 1100 MHz	The power cap triggers once SMs beyond 80 are enabled	Undervolt first; hardware mods are untested	Power
Crashes, gibberish, NaN or repeated tokens after tuning	Undervolt too deep, HBM divider (NDIV) too high, or a card that heat-soaks	Return to stock, then step back up per card; keep ECC on	Instability under load
Garbage or endless repetition with stock settings	Often an engine or recipe bug, sometimes one faulty card	Update the recipe; swap cards in `CUDA_VISIBLE_DEVICES` to find a bad one	Engine bugs
Memory test errors	A bad HBM region (fixed addresses) or a load-transient fault (errors follow timing)	Mask fixed-address regions in software; return a card with load-transient errors	Memory errors and ECC
ECC shows "Enabled" but errors still get through	Some ECC status values are filled in by the patch; whether DRAM errors are corrected is disputed	Treat ECC as a crash canary, not as protection	ECC
170tune preflight: "FBPA privilege masks are closed", `0xBADF2010`	The unlock build in use does not open the register window 170tune needs	Read the card's fuse map first, then use a matching unlock and 170tune pair	170tune
P2P says OK but transfers are wrong, or start-up hangs at 100% GPU	P2P cannot route on this topology, or the override did not apply (GNS)	Run a content check; try with P2P off	P2P
RIR tool: "debugctrl rwl bit cannot be latched"	Missing `iomem=relaxed` or the patched driver, or the card is still bound	See the prerequisites; read before you write	RIR tool
Card above 85 C, or hot even at idle	Not enough airflow; the card has no fan of its own	Forced airflow on every card, always	Cooling and power, Heat
Beeping or chirping under load	Coil noise; not a dying card	Nothing; undervolting may reduce it	Heat
PSU shuts down or the host reboots under load	An overloaded cable, a PSU with a lazy fan curve, aged breakout boards	Check cabling and PSU temperature	Power and cabling warnings
A connector or cable melted, or a PSU burned	One cable feeding two inputs, 6-pin adapters, an under-rated PSU	Stop using that setup	Power and cabling warnings
Host will not boot or shut down after unlock work	The patched driver wedges the card at every boot	Boot once with the driver blacklisted, or with the card out	Stuck host
Proxmox VM: `RmInitAdapter` failures or "WPR2 already up"	OVMF firmware, or GSP state not restored before the VM start	SeaBIOS; unlock on the host	VMs
`cuInit` returns 999, or "No CUDA GPUs are available" while nvidia-smi looks fine	Zombie CUDA processes after `kill -9`, or a container crash loop holding the card	Reboot the host, or remove the looping container	Other errors
`krcWatchdog: GPU is probably locked`, processes stuck in `uvm_va_space_destroy`, nvidia-smi hangs (in a VM)	A driver wedge, for example 615.71.09 with RTX PRO 6000 cards sharing the driver	Do not run more nvidia-smi; restore known-good modules on disk, then hard-reset the VM	VM wedge
A card stops booting (GSP booter `0xffff`) after installing an HBM clock tool	The tool writes FBPA0 PLL registers and the card has FBPA0 fused off	Switch to a driver lineage without PLL writes; read the fuse map before tuning	Warning
ESXi guest: NVML shows a 256 MiB BAR1 while sysfs shows 64 GiB	Every VM power-on resets the card's ReBAR override	A guest-side repair hook before the driver loads	ESXi
vLLM takes about 3 minutes to start serving	Normal on PCIe Gen2 links	Nothing	Source: @c133, 2026-10-09 Source: @mountains1605, 2026-10-09
Decode slower after upgrading the unlock	Unresolved (one report, v0.4 to v0.5)	Compare the patch sets; report your numbers	After an update

Combinations that failed for someone are collected on Known bad combinations.

Warnings at a glance

Warning RIR burns on 10 GB cards are permanent, a wrong address can brick the card, and 8 GB cards have no free RIR slots. Details
Warning Driver 615.78.08 breaks the unlock; use 615.71.09. Details
Warning One PCIe cable feeding a card burned a PSU; a 6-pin to 8-pin adapter melted at 170 W; a refurbished server PSU blew up at about 55% load. Details
Warning The 300 W VBIOS has burned a card. Details
Warning The card is passive: with no airflow it drops off within minutes even at idle. Details
Warning HBM clock tools that write the FBPA0 PLL stopped two cards with FBPA0 fused off from booting; read the fuse map first. Details
Warning BAR1 P2P across root ports with ACS redirect on lost a whole copy silently. Details
Warning The "CMP-170HX-64GB-Unlocked-VBIOS" repos are bogus; do not flash them. Details
Warning Cheap "new" 170HX cards from China are PG199-origin boards that unlock to 32 GB only. Details

Buying and seller warnings

More on Buying. Members flag sellers in the Discord's supplier channel, so search there before you pay. Sellers are named below as the posts named them.

Warning Crypto-only or wire-only payment. A listing that asks you to message the seller elsewhere and pay only in crypto is a scam; nobody sells these cards at 70% off. An AliExpress listing (item 3256812785475872) at USD 700 asked for crypto payment only. The Alibaba listing "LEGIT DEAL For New CMP 170HX" at USD 850 offered no Secure Pay, only wire transfer; the seller swore it was legit, called itself a bulk business partner and blocked the buyer after some discussion. Cheap 170HX listings in shopping search results are pay-by-crypto scams. A large share of requests to deal off-platform on Xianyu come from scammers.
Source: @ob88594, 2026-10-06
Source: @forte9000, 2026-10-07
Source: @domrockt2, 2026-10-07
Source: @ybryany, 2026-10-14
Source: @galygious, 2026-10-14
Source: @matstatman, 2026-10-27
Warning Marketplace search results. Asked whether AliExpress results for the 170HX are all scams: "always has been, always will be"; buy where you have buyer protection.
Source: @djpandaman, 2026-10-12
Warning Fake tracking. An Alibaba seller built a fake tracking website and tracking number and dragged out the marketplace dispute twice; the buyer got his money back only through a credit-card chargeback. Pay with a method your bank can reverse.
Source: @djpandaman, 2026-10-07
Warning Cancellations and price changes. eBay sellers cancelled orders as "out of stock" when prices rose. A seller raised the price twice in the middle of an order.
Source: @sojun80, 2026-10-07
Source: @zaknafar, 2026-10-10
Warning Payment limbo. The seller behind the Alibaba listing "170HX-B300-Turbo-GDDR6X-384Bit" shipped one card, then sat on an 8-card second order: after a T/T bank transfer it stayed in "supplier needs to confirm payment" for two weeks although the money arrived the next day. The buyer knows how to get a refund for an order that never ships, not for one stuck in that state; he also says 48 cards he had locked in earlier never shipped. Another member found the same seller appears to run several accounts.
Source: @fnerdman, 2026-10-18
Source: @adusts, 2026-10-18
Warning AI-generated product photos on a listing from a seller with nine years on the platform; ask for real photos.
Source: @con3po7016, 2026-10-14
Source: @endsjeremy, 2026-10-14
Warning Impossible unlock claims. A listing claimed the card unlocks ECC, PCIe 4.0, NVLink and 80 GB. Only ECC is in the public unlock.
Source: @lia_oc, 2026-10-13
Source: @tugoubxs, 2026-10-14
Warning Cards that cannot be tested before shipping. The Superbuy purchasing agent has no test bench for GPUs; CMP cards one member bought through Superbuy and Goofish arrived dead ("dont buy cmps using superbuy u get bricks shipped"). For hardware that can be checked by switching it on, the agent seemed reliable.
Source: @kellertreppe123, 2026-10-25
Warning Accessories. Every E1.S drive cage listing on several Chinese marketplaces asked for more money after purchase or claimed the item was broken.
Source: @cute_sc2, 2026-10-26
Warning Alibaba sellers reported by name. A buyer of 10 cards on Alibaba listed the sellers that failed him: Phat Trading Company / CONG TY TNHH DUC ("blatant scammers", profile since removed from Alibaba); Shenzhen Yige Gangbeng Technology Co., Ltd. (sat on the money for a couple of weeks, then said it could not get stock and refunded); Xi'an Qujiang New District Reicha Microelectronics Product Sales Department (tried to sell him 32 GB stock, delays, every excuse to cancel); Shenzhen Qianbao Technology Co., Ltd. ("waiting on stock", "stock was defective", delays, then a refund). Alifond was the one that delivered, and it refunded a card with bad HBM. Another member had already posted "do not buy from them" about the Xi'an Qujiang Reicha seller, with the reasons in screenshots.
Source: @kellertreppe123, 2026-10-11
Source: @boz._., 2026-10-23
Warning Sellers named only in screenshots. A seller shown in a screenshot in the supplier channel was called a scammer; he told another member he would not use trade assurance, and that member did not buy. A member says a seller trading under two accounts (also only in screenshots) still owes him about USD 2,000. Search the channel before paying a seller you have not used.
Source: @djpandaman, 2026-10-04
Source: @artistosu, 2026-10-04
Source: @spectakkular, 2026-10-04
Warning Fake delivery on "x16 unlocked" cards. A Facebook Marketplace listing, "CMP 170HX 8Gb 16x PCIe 64GB Unlock", offered cards at USD 1,200 each with a falling "left" count, and the seller talked shop convincingly and said the cards were modded locally. After three weeks of delay excuses the chat moved to email; the carrier's delivery photos showed someone else's door, and the seller would not send a photo of the label. After a PayPal "item not received" claim the seller produced a second tracking number, "delivered" to another door in the same zip code. PayPal Goods and Services on a credit card refunded the USD 3,000 the same day. A friend of another member hit the same pattern (fake UPS tracking, a seller-supplied Alibaba checkout link) and was still waiting for his card refund. Lessons from the write-up: Goods and Services on a credit card, never Friends and Family; a seller who will not show the label has answered you; save the carrier's photos the day they post; file the claim before the fake delivery scan; keep the chat with your address in writing. Full write-up on Discord.
Source: @mfj4, 2026-10-14
Source: @adusts, 2026-10-14
Before buying: dump the card's InfoROM history and test it; see Testing a card and Which card do I have.

Other warnings

Warning The Strata inference library pulls in further external libraries; one member had an agent audit it before use.
Source: @kellertreppe123, 2026-10-14
Source: @sojun80, 2026-10-14
Warning Coding agents with root or network access have left a host unreachable and edited files destructively. Details

Before you ask for help

Post these with your question; they answer most first replies.

OS and kernel (uname -r), card variant (8 GB or 10 GB), driver version, unlock fork and commit, recipe and version.
lspci -nn for the whole host and sudo lspci -vv -s <bdf> for the card.
sudo dmesg | grep -E "SEC2_DEBUG|NVRM|Xid" in full.
The newest cmpunlocker install log (logs/install_<timestamp>.log).
Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub, posted by @rh3dai, 2026-10-08
Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @rrickk420, 2026-10-09
nvidia-smi and nvidia-smi -q -d TEMPERATURE.
The card's InfoROM black box from gpcprobe's blackbox.py.
Source: repo github.com/matstatman/gpcprobe by matstatman on GitHub, posted by @matstatman, 2026-10-20
What the driver actually received: cat /proc/driver/nvidia/params and every nvidia line in /etc/modprobe.d/.
A per-boot check that has worked on an 8-card box: nvidia-smi --query-gpu=index,pci.bus_id,memory.total,pcie.link.gen.current,pcie.link.width.current,ecc.mode.current --format=csv,noheader (65536 MiB, gen 2 and the expected width on every card), nvidia-smi topo -p2p r, dmesg | grep -c Xid (0) and the P2P content check with the GPUs idle.
Source: @darkmaniac7, 2026-10-27 (lab notes)
Raw logs, not an AI summary of them: a helper could not judge a report made of formatted script output pasted from an agent, and asked what was actually run.
Source: @djpandaman, 2026-10-11
Source: @exile_1ck, 2026-10-11

Pages in this section