CMP 90HX unlock hub
More actions
The CMP 90HX is not a CMP 170HX. It is a GA102 card, the same chip family as the RTX 3080 and 3090, with 10 GB of GDDR6X on a 320-bit bus, and it is the card the cmpunlocker Discord talks about most after the 170HX. None of the work on this page is the cmpunlocker 170HX tool: it is a set of separate projects by separate people. The 90HX has no HBM, so there is no hidden memory to unlock as on the 170HX; what people unlock on it is compute throughput, graphics (GR), the power limit and the PCIe link. Every unlock below is driver-side and volatile: a full power-off re-locks the card and the unlock reapplies at the next driver load. Read a project before running it, and do not confuse a 90HX with the 10 GB 170HX (Which card do I have).
What the card is
A 3080-class GA102 card with 10 GB of GDDR6X
The 90HX is built like an RTX 3080: "3080 which cmp90 is based on has 320W". It is 2.5 slots wide, so plan the spacing; one owner's card is starved of air on an open frame. How fast it is after an unlock depends on the workload (see #How fast it is after an unlock (contested)).
Reported
Source: @emogop, 2026-10-13
Source: @lenxtroz, 2026-10-13
Source: @javicabrera_30495, 2026-10-18
How fast it is after an unlock (contested)
The reported compute level spans a wide range and depends on the workload, so read it as workload dependent, not a single number.
- On 2026-07-20 a Shenzhen Huateng sheet quoted about 105 USD for one, and @a.civardagezen and @jonathan.quai (who owns several) placed it as a 3080-class 10 GB GDDR6X card on a 320-bit bus. Source: @a.civardagezen, 2026-07-20
Source: @jonathan.quai, 2026-07-20 - @giaki3003 (2026-08-31) put it at about an RTX 3060, with no ray tracing or frame generation, stuck at PCIe 2 even with the patch, and power limited. Source: @giaki3003, 2026-08-31
- @lenxtroz (2026-08-21) reported it as fast as an RTX 5060 in an advanced ComfyUI workflow. Source: @lenxtroz, 2026-08-21
- @javicabrera_30495 (October) called an unlocked 90HX 3090-level compute that keeps its 10 GB. Source: @javicabrera_30495, 2026-10-18
Contested
PCIe: the capacitor mod and the link cap
Capacitors restore the width; a modified driver or a Gen2 project restores the generation
Like the 170HX, the 90HX ships with the capacitors on the PCIe lines missing, so it runs PCIe 2.0 x4; soldering them restores x16. @emogop had a local technician fit his. PCIe 2.0 itself comes through a modified driver; without it his card showed Code 43. The modified driver uses the usual signing, not test signing. Asked on 2026-10-16 whether PCIe 4.0 was unlocked, @giaki3003 thought the highest is still 2.0, and @ob88594 said "PCIe 3.0 cracking is not yet complete". For the 170HX version of the same mod see PCIe x16 capacitor mod.
Reported
Source: @emogop, 2026-10-13
Source: @giaki3003, 2026-10-16
Source: @ob88594, 2026-10-13
The lane capacitors: 220 nF, pads on one side, orientation differs from the 170HX
The 90HX needs only capacitors for PCIe lane restoration: 220 nF with 5% tolerance, the same value owners use on the 170HX. @ju5t5omedude, who did the mod on two 90HX cards, found the pads on one side and the teardown annoying (get the blower model if noise does not matter), and warned that the CMP 70HX and 90HX switch capacitor orientation, so follow the traces carefully. @d3m3n7r soldered the missing capacitors on his, after which the bus reported x16; he later reported very low bus load and slightly higher, steadier FPS.
Reported
Source: @green_dam, 2026-08-28
Source: @ju5t5omedude, 2026-09-01
Source: @d3m3n7r, 2026-09-07
With the caps fitted plus a patch: Gen2 x16, DLSS, but not mandatory ray tracing (2026-09-07)
With the caps fitted and @green_dam's patch, @emogop got PCIe Gen2 x16 on 2026-09-07; DLSS worked but a game that requires hardware ray tracing would not launch at that point. The link tops out at Gen2: @giaki3003 confirmed the highest PCIe generation is still 2.0.
Reported
Source: @emogop, 2026-09-07
Source: @green_dam, 2026-09-07
Source: @giaki3003, 2026-10-16
Power, idle and VRAM cooling
pstated brings idle from about 80 W to about 7 W
A 90HX idles at about 80 W with the memory pinned at full speed; the pstated daemon brings it to about 7 W. @emogop confirmed it helped. It does not work on the 170HX (see P-states and idle power). On 2026-09-01, asked for the new idle, @serplay228 answered "12,7" (unit not given, likely watts) and said the real problem with that machine was random shutdowns from a botched mod, not pstated.
Proven
Source: @emogop, 2026-08-14
Source: @ju5t5omedude, 2026-08-14
Source: @serplay228, 2026-09-01
VRAM temperatures, and a metal plate instead of pads
With the stock pads the GDDR6X easily goes past 100 C. @emogop replaced the thermal pads on the VRAM with a copper or aluminium plate, with no insulation layer, and calls the plate mod and the capacitor mod both a must on a 90HX. The plate cost about 9 USD on the Russian AliExpress site; @lenxtroz found the same part for about 20 USD on the global site.
Reported
Source: @emogop, 2026-10-13
Source: @lenxtroz, 2026-10-13
The power limit and a V/F undervolt (850 to 875 mV)
The modified driver also raises the power limit; the ceiling is about 300 or 320 W, "apparently it's a hardware limit of that card". Unlike the CMP 40HX, the 90HX exposes a normal desktop V/F curve in MSI Afterburner. @emogop's advice is to find the maximum clock at 850 to 875 mV, since above that it heats up like all Ampere cards, and because the card is power-starved at 250 W (against 320 W on a 3080) with 60 to 70 W going to the GDDR6X, benching around RTX 3060 Ti level. @d3m3n7r applied a curve, passed a PhysX bench and saw lower temperatures.
Reported
Source: @emogop, 2026-09-17
Source: @d3m3n7r, 2026-09-17
Source: @wakeupmbrc1, 2026-09-17
Memory and SMs
No memory to unlock; 20 GB needs a physical chip or board swap
The 90HX has no memory to unlock; the only route to more is a manual upgrade with newer memory packages. For 20 GB you need a 3090 PCB plus ten more GDDR6X chips soldered on, the same as the RTX 3080 20 GB mod. @da6b asked and @programmerboi confirmed there is physically no additional memory on the board. Earlier ideas (soldering 2 GB chips, repositioning configuration straps, a modded VBIOS) were discussed but none was shown working, and @xeax put the 24 GB route at a clamshell 3090-type PCB with 20 chips.
Theory
Source: @programmerboi, 2026-10-25
Source: @xeax, 2026-10-18
Source: @da6b, 2026-10-25
No SM unlock shown; GA102 has no RIR
No SM unlock for the 90HX has been shown. Asked whether it could reach RTX 3080 SM counts, @scallion_9883 answered that GA102 and later consumer Ampere has no RIR (it is "a holdover from turing"), so an unlock would have to go through the floorsweep path, the same position as the 170HX 8 GB card; @matstatman suggested first looking for disabled GPCs or TPCs. @emogop was told his card has no RIR records. An AI-written comparison table claimed a 90HX can be modified to all 84 SMs; @emogop had seen no such release, and the table's author said that row came from his AI and may be wrong. See Full SM unlock for how the 170HX does it.
Theory
Source: @emogop, 2026-10-17
Source: @matstatman, 2026-10-17
Source: @scallion_9883, 2026-10-17
Source: @emogop, 2026-10-18
Tensor rate: the consumer FP32-accumulate cap, and the "free 3090 speedup"
The 90HX carries the same consumer tensor-rate limit as other GA10x cards: the speed-select fuses that cap FP16/BF16 math with FP32 accumulate read a throttled value on the 90HX against the full value on an RTX 3090. The Reddit "free 3090 speedup hack" (shared on the server, verified for about 15% prefill and 5% decode) works on a 90HX too; @lenxtroz found he also had to unplug the machine from the wall for it to take effect. The register-level background is on Other CMP cards and other NVIDIA GPUs.
Reported
Source: @green_dam, 2026-10-20
Source: @68x0, 2026-10-22
Source: @lenxtroz, 2026-10-22
Compute unlock projects and how to run them
These are separate 90HX projects, not the cmpunlocker 170HX tool. Each restores SM compute throughput that the card ships throttled. All are driver-side and volatile: a full power-off re-locks the card, and PCIe stays at the card's hardware link cap unless a separate Gen2 project is applied. Read a project before running it.
bendy2/cmp90hx: persistent compute unlock on NVIDIA Open 580.159.03
bendy2/cmp90hx is a persistent compute unlock for the 90HX (GA102) on NVIDIA Open 580.159.03. It uses a temporary bootstrap driver during boot and hands control back to the untouched stock driver, with a systemd service that re-applies the unlock at every boot (about 2 minutes for one card, about 11 minutes for a verified six-card host). It does not touch memory, ECC, VBIOS or PCIe link speed. Install with git clone then sudo ./install.sh and reboot; confirm with systemctl status cmp90hx-persistent.service (expect active) and the read-only check.sh, which reports the nine SM issue-rate fields as full (an INCONCLUSIVE field means the query timed out, not that the card is still limited). Do not run nvidia-smi, miners or CUDA until the batch status file shows every card done. Uninstall with sudo ./remove.sh --yes then reboot. The README is Chinese only and there is no licence file of its own (it vendors NVIDIA's open-gpu-kernel-modules source under NVIDIA's COPYING). @ju5t5omedude shared it on 2026-07-31 and @_j3ns welcomed people working on other cards.
Reported
Source: repo github.com/bendy2/cmp90hx by bendy2 on GitHub, posted by @ju5t5omedude, 2026-07-31
Source: repo github.com/bendy2/cmp90hx by bendy2 on GitHub, posted by @giaki3003, 2026-08-20
WildFlash1st/cmp90hx-unlock: compute unlock baked into NVIDIA Open 610.43.03
WildFlash1st/cmp90hx-unlock takes the compute unlock to NVIDIA Open 610.43.03 and bakes it into the driver (no systemd service, survives reboots), reusing patches from pearlfortune's 90HX work. Install with sudo ./install.sh --profile=cmp90-rejoin15 after staging the stock 610.43.03 userspace and GSP firmware, then reboot and verify with check.sh (nine issue-rate fields full). Its own measurements: prefill about 224 to 1770 tok/s on a 12B Q4 model after the unlock. PCIe stays Gen1 x16, which the repo attributes to a hardware link cap of the card (LnkCap2 2.5 GT/s), not a driver limit; NVreg_EnablePCIeGen3=1 changes nothing. A separate cmp90hx-unlock-for-windows does the Windows side. Rollback with sudo ./remove.sh --yes and reboot.
Reported
Source: repo github.com/WildFlash1st/cmp90hx-unlock by WildFlash1st on GitHub, posted by @68x0, 2026-09-04
Source: repo github.com/WildFlash1st/cmp90hx-unlock by WildFlash1st on GitHub, posted by @emogop, 2026-08-31
pearlfortune/cmpunlocker: a Chinese CLI for the 170HX, 90HX and 50HX
pearlfortune/cmpunlocker (often "the Chinese unlock", unrelated to the amoghmunikote tool despite the name) is a prebuilt CLI that removes the compute limiter on the CMP 170HX 8 GB, 90HX and 50HX you own, Linux x86-64, root required. It does not flash the VBIOS and refuses to run if the driver, kernel or VBIOS is not in its tested table. For the 90HX it offers a temporary unlock (lost on reboot, just run it again) and a 610.43.03 persistent engineering preview; success prints PASS_CMP90HX_ALL_TARGETS_FULL_SPEED. @funzr ran one 90HX plus one 50HX, both power-limited to 140 W, on it: Qwen 3.6 35B A3B Q4 at 128k context, about 1700 tok/s prefill and 75 tok/s decode. Download from the project's GitHub releases and verify the SHA256SUMS before extracting.
Reported
Source: repo github.com/pearlfortune/cmpunlocker by pearlfortune on GitHub, posted by @adih1266, 2026-08-04
Source: repo github.com/pearlfortune/cmpunlocker by pearlfortune on GitHub, posted by @funzr, 2026-08-19
jdowning100/cmpunlocker: adds a PCIe Gen2 step on top of the compute unlock
jdowning100/cmpunlocker adds PCIe Gen2 (5.0 GT/s) on top of the pearlfortune compute unlock (the "rejoin16" notes), ending at a Gen2 link with compute still full speed; its apply step runs a series of driver-reload cycles (about 7 to 8 minutes) and retrains the link. Its 170HX part is a stale August snapshot: use mainline for the 170HX. The repo carries a host-crash warning (see below).
Reported
Source: repo github.com/jdowning100/cmpunlocker by jdowning100 on GitHub, posted by @mikevlz_59313, 2026-09-04
Confirmed unlocked, about 90 TFLOP/s FP16/BF16
@weicj reported the 90HX fully compute-unlocked at about 90 TFLOP/s FP16/BF16 on 2026-08-12, and @serplay228 called the pure performance unlock confirmed on 2026-08-18, with much better AI performance, while other 90HX claims lacked confirmation at the time.
Reported
Source: @tekfabii, 2026-08-18
Source: @serplay228, 2026-08-18
Whether FP32 can be software-unlocked (contested)
- WildFlash1st's analysis states software unlock of FP32 on the 90HX is impossible. Source: @emogop, 2026-08-20
- jdowning100's benchmarks show some FP32 operations working. Source: @giaki3003, 2026-08-20
The two were not reconciled in the export.
Contested
Six pooled 90HX cards against one unlocked 170HX
@ju5t5omedude pooled six compute-unlocked 90HX cards and found them only marginally slower than an unlocked 170HX in his benchmark; no workload, numbers or memory-bound comparison were given, so the scope is unclear.
Reported
Source: @ju5t5omedude, 2026-09-01
Graphics and gaming
Icyoung/cmpunlocker tools/gr_unlock: enable the Graphics engine
Icyoung/cmpunlocker tools/gr_unlock enables the Graphics (GR) engine (3D, display output, OpenGL, Vulkan and CUDA graphics) on the CMP 90HX (10de:220d) and CMP 70HX (248a) by patching NVIDIA's closed-source 610.43.02 driver, following the GreenDamTan lineage. Run sudo ./install_90hx.sh (it fetches the stock 610.43.02 driver, patches it, installs a helper and prompts to reboot); a boot service then applies the hardware pass. It is a separate, closed-driver install that cannot coexist with the 170HX open-driver path on the same host, and it does not restore NVENC, RT or persistent Gen2. Reboot rather than reload the module, and it still needs a second GPU or an iGPU to light a monitor. Remove with sudo ./remove_90hx.sh.
Reported
Source: repo github.com/Icyoung/cmpunlocker by Icyoung on GitHub
The graphics and Gen2 unlocks originate outside this server
The 90HX Gen2 and graphics unlocks originate outside the cmpunlocker server, on an external file share (a UEFI loader plus @green_dam's patched Windows driver, packaged by another user), and work in EFI on Windows and Linux. The patched driver raises GR and PCIe only. On Linux @green_dam confirmed Gen2 working (2026-08-28). @jonathan.quai opened a draft 90HX PR against cmpunlocker; the maintainer recommended a separate fork rather than folding it in. @emogop also pointed to a WildFlash1st Windows EFI unlock that unlocks compute though games still run badly, and @giaki3003 summarised the Windows recipe as the EFI unlock for the features plus the patched driver for PCIe.
Reported
Source: repo github.com/WildFlash1st/cmp90hx-unlock by WildFlash1st on GitHub, posted by @green_dam, 2026-08-28
Source: @68x0, 2026-08-28
Source: @giaki3003, 2026-08-28
Source: @jonathan.quai, 2026-08-14
A Windows power-limit unlock driver
On 2026-09-26 @emogop reported a newer Windows unlock plus driver that raises the power limit; with it nvidia-smi showed the card at PCIe Gen2 x16 and a 300 W limit set with nvidia-smi -pl 300, and @green_dam advised adjusting the voltage so the extra power can actually be drawn. @emogop said the card then games like a 40HX used to before its unlock, only faster.
Reported
Source: @emogop, 2026-09-26
Source: @green_dam, 2026-09-26
Ray tracing: not working, then working (Indiana Jones runs)
The ray-tracing state changed over September and October. On 2026-09-07 a game with mandatory ray tracing would not launch; on 2026-09-26 the newest mod driver advertised RT cores and crashed ray-traced games (@d3m3n7r confirmed anything using RT crashes); on 2026-10-16 @emogop reported hardware RT working, with a screenshot of Indiana Jones (a game that needs hardware RT) running, crediting another member and friends. The October result supersedes the September dispute. An earlier benchmark had shown RT not usable: @emogop measured Quake 2 RTX at 25 to 30 fps against 60 fps on an RTX 3070M.
Reported
Source: @emogop, 2026-09-13
Source: @d3m3n7r, 2026-09-26
Source: @emogop, 2026-10-16
Measured gaming
On an unlocked 90HX, @d3m3n7r got Metro Exodus about 70 to 90 fps and GTA 40 to 70 fps; @emogop got FF7 Rebirth to a solid 60 fps at 4K with DLSS Quality after a community DX12 fix. On the dartraiden drivers alone (permission to use the card for 3D only, no performance) it was far slower, about 9 fps in Unigine Heaven. The dartraiden patch is a closed-driver patch, so it cannot be combined with an open-module compute unlock.
Reported
Source: @d3m3n7r, 2026-09-07
Source: @emogop, 2026-09-15
Source: @serplay228, 2026-08-18
CUDA support (contested)
- @ob88594: "It seems CUDA calls are not supported and will cause errors", and later that performance calls are ineffective and AI performance is much worse than a 3080. Source: @ob88594, 2026-10-13
- @xnupanic76: "The 90HX works fine for CUDA, I have one in one of my workstations." Source: @xnupanic76, 2026-10-13
The compute unlocks above (bendy2, WildFlash1st, pearlfortune) raise the SM issue rates and were verified on multi-card hosts, which supports @xnupanic76, but nobody posted a CUDA sample on the stock driver, so this stays contested.
Contested
ECC, NVLink and GSP
ECC enable attempt (unconfirmed)
@green_dam tried to enable ECC on a 90HX by patching the driver's ECC functions; the driver showed ECC enabled, but he could not confirm it works, and members of a Chinese group doubted it. He noted that enabling ECC on a Tesla P4 usually reduces usable memory.
Theory
Source: @green_dam, 2026-09-05
Source: @cmurrayis, 2026-09-05
Source: @skyflyingfish, 2026-09-08
No NVLink; GSP-off keeps the unlock across a driver reload
The 90HX has no NVLink fingers ("no splines"), and a bridge would cost more than the card anyway. The GSP-off approach @zarknoth described for CMP cards (unlock first, then load the driver with GSP firmware disabled) "work well" on the 90HX, and the non-GSP driver can be modified to use vGPU (screenshot, single report). The 170HX side of this is on VBIOS and firmware.
Reported
Source: @minioadmin, 2026-10-19
Source: @scallion_9883, 2026-10-19
Source: @zarknoth, 2026-10-27
Source: @green_dam, 2026-10-28
PCIe Gen2 troubleshooting and warnings
Compute unlocked, but the Gen2 step leaves the link at Gen1
@mikevlz got the compute unlock on 610.43.03 but the Gen2 (rejoin16) step left PCIe at Gen1; @jonathan.quai pushed commits for a suspected cold-boot timing issue. Users must compile the small bar0 helper tools themselves or the scripts do not run. @olehborzykh hit the same thing in a combined build until the Gen2 patch was actually merged into it (see #In a Proxmox or Windows VM).
Reported
Source: @mikevlz_59313, 2026-08-15
Source: @jonathan.quai, 2026-08-16
The Windows driver forces Gen1; a registry helper caused a blue screen
Warning
On Windows the 90HX driver forces the link to Gen1 at every load, so setting the link speed from outside does not stick. @giaki3003's CMP90HX-Gen2-Helper edits the display-class registry key; its option 2 caused a low-resolution blue screen right after the Windows logo, recovered with a System Restore point. Editing these registry values is dangerous. @green_dam concluded the Windows driver itself needs patching. Do not edit these keys without a restore point.
Source: @giaki3003, 2026-08-24
Source: @emogop, 2026-08-24
Source: @green_dam, 2026-08-24
Letting the driver negotiate Gen3 floods AER errors
Letting the driver negotiate PCIe Gen3 on its own after register edits flooded the logs with AER errors. Do not force Gen3 on the 90HX; its link tops out at Gen2.
Reported
Source: @mishanya13, 2026-08-30
Source: @green_dam, 2026-08-30
Source: @giaki3003, 2026-08-30
The Chinese graphics unlock drops the card off the bus on Linux
@serplay228 tried the 90HX graphics unlock from the external share on Linux; the driver failed to initialise with the card fallen off the bus. He later said he had likely broken something and that it would not work normally with that unlock. @green_dam clarified that unlock is applied in EFI and the patched Windows driver only raises GR and PCIe.
Reported
Source: @serplay228, 2026-08-31
Source: @green_dam, 2026-08-31
Host-crash warning on the Gen2 (rejoin16) path
Warning
The jdowning100 repo carries a host-crash warning: a botched PCIe link write can wedge the GPU until reboot, and on a host where the GPU shares a PCIe switch with the NIC or SATA a wedged GPU can take the whole host down. Do not apply a Gen2 link write on a card that shares its PCIe switch with storage or network functions.
Source: repo github.com/jdowning100/cmpunlocker by jdowning100 on GitHub, posted by @mikevlz_59313, 2026-09-04
Running models
At PCIe Gen1 x4, llama.cpp prefill is the bottleneck
With the 90HX compute-unlocked but PCIe still at Gen1 1.1, @mikevlz_59313 (2026-08-18) gets only 50 to 70 tok/s prefill in llama.cpp on Qwen3.6-35B-A3B at Q4, while decode is about 30 tok/s; nvtop shows PCIe saturated at about 860 MB/s. He runs his own FMA and dp4a substitution patches on vanilla llama.cpp. @ju5t5omedude pointed to llama.cpp patches that work on the 90HX: Arabella's (ggml-org llama.cpp issue 24616) and eastmoe's auto-patch script. Models that fit in VRAM run smoothly; MoE models that do not fit are where prefill hurts.
Reported
Source: @mikevlz_59313, 2026-08-18
Source: @mikevlz_59313, 2026-07-27
Source: @ju5t5omedude, 2026-07-27
On an old platform, and as a TTS/ASR sidecar
In a Proxmox or Windows VM
A Windows gaming VM at Gen2 x16
@icych runs 90HX cards at Gen2 x16 under Arch Linux with Windows in QEMU via GPU passthrough; performance is about RTX 3060 level, and Stalker 2, Atomic Heart and Helldivers at medium settings in FHD with a 200 W power limit gave 70 to 90 fps. @emogop's card at PCIe 1.1 x4 performed about like a GTX 1060. The 90HX is heavily power-limit bound: its GDDR6X draws about 80 W at max clocks, leaving about 120 W for the GA102 at a 200 W cap, so undervolting helps.
Reported
Source: @icych, 2026-08-31
Source: @emogop, 2026-08-31
A combined 170HX + 90HX driver on one Proxmox host
@olehborzykh runs a 170HX 40 GB and a 90HX in one Debian 13 / Proxmox host with a custom combined build of NVIDIA Open 610.43.03. A Gen2 watcher never saw a link flip until the Gen2 patch was actually merged into the combined build (diagnosed by @kmike84); after merging it and rebooting, the link retrained to Gen2 x4 and both the 170HX 40 GB unlock and the 90HX compute unlock still worked. He said the unlock is plain cmpunlocker and his problem was the patch application steps. A longer combined-driver write-up by @olehborzykh exists; see Discord.
Reported
Source: @olehborzykh, 2026-09-01
Source: @kmike84, 2026-09-01
SPI flash failure (warning)
Warning
A 90HX @d3m3n7r bought for about 78 USD showed "GOP header is not present" at boot and Code 43 in Windows, with every BIOS write failing. On teardown an SPI flash leg was lifted; after he resoldered it the firmware flashed on the first try, but the flash then periodically lost its VBIOS and had to be reflashed, which he put down to thermal degradation from mining on the stock cooler. On 2026-09-18 he confirmed the flash is dying: a dump read after a VBIOS loss differed from the freshly flashed image. If a mining card loses its VBIOS after a reflash, suspect a dying SPI flash chip, not the unlock. This is very likely the same card as the Code 43 / "DQ memory training failed" case once filed under the 170HX (Troubleshooting): a discrete memory chip and an EFI unlock point to a 90HX, not a 170HX (whose HBM is on package).
Source: @d3m3n7r, 2026-09-06
Source: @ins_mniac, 2026-09-06
Source: @d3m3n7r, 2026-09-17
Source: @emogop, 2026-09-17
Price and whether it is worth it
Cheap where mining cards are cheap, about an RTX 3080 elsewhere
@d3m3n7r bought a 90HX in Ukraine for about 78 USD, against about 315 USD for an RTX 3080. @wakeupmbrc1 saw them on Avito at about 50 to 100 USD. @roboalchemist. said eBay prices are roughly 1:1 with RTX 3080 prices, and @bayley_w said the 90HX is at e-waste prices outside the US. @ju5t5omedude paid about 140 USD each when only the compute unlock existed and said he should have bought more 170HX cards instead. By late October a Huateng sheet listed the 90HX at 248 USD. The cards often arrive looking nearly unused, with fresh paste. Since a 90HX now costs about as much as an RTX 3080, there is little point unless you already own one. More prices on Price history and market.
Reported
Source: @d3m3n7r, 2026-08-30
Source: @roboalchemist., 2026-08-31
Source: @wakeupmbrc1, 2026-08-31
Source: @bayley_w, 2026-09-01
Source: @ju5t5omedude, 2026-09-01
Source: @legitsplit, 2026-10-26
Mislabelled "170HX" listings are often a 90HX or a weaker card
Warning
Many Alibaba listings titled CMP 170HX are wrong: on 2026-07-21 @meatsus saw 170HX listings whose purchase options were "90HX 12GB" and "90HX 20GB", and nobody could find a listing that gave the memory correctly (they all said GDDR6; the 170HX uses HBM2e). One zero-feedback seller's listing was in fact a CMP 30HX, a much weaker card. Ask for proof of testing before paying, and treat listing photos as no evidence of real stock.
Source: @skirtchaser, 2026-07-21
Source: @meatsus, 2026-07-21
Source: @giaki3003, 2026-07-21
Scope: cmpunlocker stays 170HX-only
Asked repeatedly to add the 90HX, 100HX and other CMPs, the cmpunlocker maintainer (@_j3ns) kept the tool to the 170HX: on 2026-08-14 "I might consider it, but only after I finish the 170HX", and on 2026-09-14 "I'm going to stick to 170HX only", because maintaining it is already a lot of work. @bayley_w called the 90HX much less interesting ("all that effort gets you a 3080"); @djpandaman noted that work on unlocks that apply to both cards happens in the "clean room", not the main tool. Everything above is separate projects by separate people.
Reported
Source: @mrmannyatx, 2026-07-31
Source: @bayley_w, 2026-08-06
Source: @_j3ns, 2026-08-14
Source: @_j3ns, 2026-09-14
Source: @djpandaman, 2026-09-14