Tools - Revision history

ImportBot: Import: tools overview, @mountains1605, @globalapathy, @xeax, @matstatman, @c133, @wakeupmbrc1, @zagitta, @impactic, @djpandaman, @sojun80, @arabus_74064, @reala7096, @_arcanum_, @rh3dai; repos PixelML/club-170hx, SeraphimSerapis/tool-eval-bench, dev-boz/cmp170hx-testsuite, lixiangwuxian/170hx-rir-tool, sasha0552/nvidia-pstated, zzc0721/torch-performance-test-data; gap-fill from #general 2026-10-23 to 10-25: @phoenixone7, @zagitta, @mrwojaketh, @danske1, @petrikrohn_26497, @d

2026-10-30T10:02:39Z

Import: tools overview, @mountains1605, @globalapathy, @xeax, @matstatman, @c133, @wakeupmbrc1, @zagitta, @impactic, @djpandaman, @sojun80, @arabus_74064, @reala7096, @_arcanum_, @rh3dai; repos PixelML/club-170hx, SeraphimSerapis/tool-eval-bench, dev-boz/cmp170hx-testsuite, lixiangwuxian/170hx-rir-tool, sasha0552/nvidia-pstated, zzc0721/torch-performance-test-data; gap-fill from #general 2026-10-23 to 10-25: @phoenixone7, @zagitta, @mrwojaketh, @danske1, @petrikrohn_26497, @d

← Older revision		Revision as of 10:02, 30 October 2026
Line 85:		Line 85:
	* {{Status\|Reported}} '''vLLM run-to-run variance.''' Re-running the same quantization comparison gave results about 3 percent apart; @vektor_prime traced it to upstream vLLM and fixed it with a batch-invariance change in his recipe repository. {{Source\|handle=vektor_prime\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1557467960455467170}}<br />{{Source\|handle=vektor_prime\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1557498798530498601}}		* {{Status\|Reported}} '''vLLM run-to-run variance.''' Re-running the same quantization comparison gave results about 3 percent apart; @vektor_prime traced it to upstream vLLM and fixed it with a batch-invariance change in his recipe repository. {{Source\|handle=vektor_prime\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1557467960455467170}}<br />{{Source\|handle=vektor_prime\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1557498798530498601}}
	* {{Status\|Contested}} '''LMCache''' for spilling KV cache to CPU RAM: with it vLLM processed one request at a time instead of eight, and @c133 found it also wanted VRAM. {{Source\|handle=nikhilr1696\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1556742092150284339}}<br />{{Source\|handle=c133\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1556741812394270760}}		* {{Status\|Contested}} '''LMCache''' for spilling KV cache to CPU RAM: with it vLLM processed one request at a time instead of eight, and @c133 found it also wanted VRAM. {{Source\|handle=nikhilr1696\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1556742092150284339}}<br />{{Source\|handle=c133\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1556741812394270760}}
			* {{Status\|Reported}} '''NeuralFree''' ([https://github.com/EonMux/NeuralFree EonMux/NeuralFree]): a work-in-progress inference framework by @phoenixone7. GGUF support is in; safetensors support and testing are still to come. He split it out of an older trading-bot project and plans to keep supporting it because of prefix-caching problems he hit in vLLM. No CMP numbers posted. {{Source\|handle=phoenixone7\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563213261850157077\|repo=https://github.com/EonMux/NeuralFree}}<br />{{Source\|handle=phoenixone7\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563213796913188956}}
	* {{Status\|Theory}} '''EAGLE-4''' speculative decoding was flagged as faster than MTP; nobody tried it on a 170HX. {{Source\|handle=cibernox\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562247115562942555}}		* {{Status\|Theory}} '''EAGLE-4''' speculative decoding was flagged as faster than MTP; nobody tried it on a 170HX. {{Source\|handle=cibernox\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562247115562942555}}
	* {{Status\|Theory}} '''flare-local-inference''' (NVIDIA) is built for Blackwell; a port to these cards would need a purpose-written engine. {{Source\|handle=mountains1605\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562492007262322789}}<br />{{Source\|handle=dufrelic\|date=2026-10-22\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562824922286526505}}		* {{Status\|Theory}} '''flare-local-inference''' (NVIDIA) is built for Blackwell; a port to these cards would need a purpose-written engine. {{Source\|handle=mountains1605\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562492007262322789}}<br />{{Source\|handle=dufrelic\|date=2026-10-22\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562824922286526505}}
Line 96:		Line 97:
	* Iron: a Pinecil or any 60 W or stronger iron with SMD tips; the Sequre S99 (USB-C, C245 tips, 100 W). {{Source\|handle=civiloid\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558803228362145833}}<br />{{Source\|handle=globalapathy\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558832357514154055}}		* Iron: a Pinecil or any 60 W or stronger iron with SMD tips; the Sequre S99 (USB-C, C245 tips, 100 W). {{Source\|handle=civiloid\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558803228362145833}}<br />{{Source\|handle=globalapathy\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558832357514154055}}
	* Microsoldering tweezers are essential; the capacitors are almost impossible to handle otherwise. Clean with isopropyl alcohol, foam swabs, a brush and wipes. {{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558857956961820794}}<br />{{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558858884343668859}}		* Microsoldering tweezers are essential; the capacitors are almost impossible to handle otherwise. Clean with isopropyl alcohol, foam swabs, a brush and wipes. {{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558857956961820794}}<br />{{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558858884343668859}}
	* Microscope: @civiloid would rather have a QianLi Gorilla 6558 or Mechanic 6560T with a camera than work off a screen alone. {{Source\|handle=civiloid\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558803596852596837}} A used Vision Engineering TS-4 is another option; skip the TS-3. {{Source\|handle=cmdrwaka324\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559782730688893020}}		* Microscope: @civiloid would rather have a QianLi Gorilla 6558 or Mechanic Robot 6560T with a camera than work off a screen alone. {{Source\|handle=civiloid\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558803596852596837}} For soldering he calls a new Chinese binocular microscope with a camera module the most cost-effective choice: small, good enough optics, and the camera plugs into any screen. QianLi and Mechanic are consumer brands that buy white-label units, so the actual manufacturer is cheaper. {{Source\|handle=civiloid\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559808745176043613}}<br />{{Source\|handle=civiloid\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559808927581872161}}<br />{{Source\|handle=civiloid\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559809944977408102}} @_arcanum_ paid about USD 300 on Taobao for the QianLi 6558 UV-light version; the IR module is about 140 more. {{Source\|handle=_arcanum_\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559983772076609640}}<br />{{Source\|handle=_arcanum_\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559985698352795670}}<br />{{Source\|handle=_arcanum_\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559985805802340454}} A used Vision Engineering TS-4 is another option; skip the TS-3, which is louder (fans on the supply and the halogen housing) and older in optics and electronics. {{Source\|handle=cmdrwaka324\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559782730688893020}}<br />{{Source\|handle=cmdrwaka324\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559798125542248489}} On the TS series, video output needs the camera optics (the CS adapter) fitted when bought; scopes in this class run USD 2,000 to 3,000, and eyepiece-only scopes cost about 5x less. Many cheap camera heads send one eye's light fully to the camera, so you lose binocular view while recording. {{Source\|handle=endsjeremy\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559786404223914037}}<br />{{Source\|handle=cmdrwaka324\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559788016095076363}}<br />{{Source\|handle=endsjeremy\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559804501127462943}}<br />{{Source\|handle=endsjeremy\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559805792067125259}}<br />{{Source\|handle=endsjeremy\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559984711525343233}}
	* A solder stencil for the mod is in the channel pins, but making one needs a laser cutter. {{Source\|handle=globalapathy\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1564479737160802307}}<br />{{Source\|handle=globalapathy\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1564479895952695377}}		* A solder stencil for the mod is in the channel pins, but making one needs a laser cutter. {{Source\|handle=globalapathy\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1564479737160802307}}<br />{{Source\|handle=globalapathy\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1564479895952695377}}
	* A video of the mod done quickly by a Canadian GPU repair YouTuber: https://www.youtube.com/watch?v=HO2T4TgNzlQ {{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558856094116483142}}		* A video of the mod done quickly by a Canadian GPU repair YouTuber: https://www.youtube.com/watch?v=HO2T4TgNzlQ {{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558856094116483142}}
Line 109:		Line 110:
	== Related work on other NVIDIA cards ==		== Related work on other NVIDIA cards ==
	* {{Status\|Reported}} '''Zagitta/gpu-unlock''': a Linux kernel module that removes the half-rate tensor-math limit on RTX 3090 class cards with the register approach the CMP unlocks use, without driver patching. It must load before the NVIDIA driver; it only matches RTX 3090 IDs by default. Gain: a few percent prefill when the engine uses BF16 to FP32 FMA; no core or SM unlock. {{Source\|handle=zagitta\|date=2026-10-18\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1561189774834802709}}<br />{{Source\|handle=zagitta\|date=2026-10-19\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1561876551505551411}}<br />{{Source\|handle=zagitta\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562214717316137005}}<br />{{Source\|handle=zagitta\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563321762756694107}} @sojun80 used the approach on a 3060 Ti, 3070, 3070 mobile, 3080 and 3080 Ti, and is building a generic unlock tool from cmpunlocker for the 30 and 40 series. {{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562322577676967957}}<br />{{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562322921383264289}} Details belong on [[Unlocks]].		* {{Status\|Reported}} '''Zagitta/gpu-unlock''': a Linux kernel module that removes the half-rate tensor-math limit on RTX 3090 class cards with the register approach the CMP unlocks use, without driver patching. It must load before the NVIDIA driver; it only matches RTX 3090 IDs by default. Gain: a few percent prefill when the engine uses BF16 to FP32 FMA; no core or SM unlock. {{Source\|handle=zagitta\|date=2026-10-18\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1561189774834802709}}<br />{{Source\|handle=zagitta\|date=2026-10-19\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1561876551505551411}}<br />{{Source\|handle=zagitta\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562214717316137005}}<br />{{Source\|handle=zagitta\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563321762756694107}} @sojun80 used the approach on a 3060 Ti, 3070, 3070 mobile, 3080 and 3080 Ti, and is building a generic unlock tool from cmpunlocker for the 30 and 40 series. {{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562322577676967957}}<br />{{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562322921383264289}} Details belong on [[Unlocks]].
			** @grapepepper patched his RTX 3090 Ti from the repository; after he shared his results others unlocked various RTX 30-series cards, and he has seen a report of an RTX 4070 Ti SUPER. {{Source\|handle=grapepepper\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562183666371924070}}
			** '''eGPU.''' Zagitta: someone got it working on an eGPU (see the repo issues). {{Status\|Theory}} @da6b, whose 3090 sits behind a Thunderbolt 3 bridge, expects to need the Thunderbolt drivers in the initramfs and Thunderbolt security disabled so the tunnel is up before the module loads, possibly with some timing involved; an eGPU on M.2 is a direct PCIe link and simpler. {{Source\|handle=zagitta\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563321076027752459}}<br />{{Source\|handle=da6b\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563320911858376785}}<br />{{Source\|handle=da6b\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563321227320492143}}
			** {{Status\|Theory}} @sojun80 thinks the tensor limit applies to the 20 to 40 series, "i think all of them"; he has no 20-series card left to test. {{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562275001879629885}}<br />{{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562322138516684843}}
			** Zagitta on its origin: it builds on the CMP unlock work (the same basis as cmpunlocker, and the CMP 90HX and 50HX unlocks), and nothing in it is new; his only change is doing it without a driver patch, as a standalone early-boot module, an idea he took from the UEFI module the 90HX and 50HX Windows unlocks use. He has no Windows version planned. On 2026-10-24, asked whether it covers only GA102, he said GA104 and GA106 as well; other GPUs in the same host (a 50-series card, for example) are left alone. {{Source\|handle=zagitta\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563146783062884393}}<br />{{Source\|handle=zagitta\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563529939586912410}}<br />{{Source\|handle=mrwojaketh\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563529790848372796\|repo=https://github.com/Zagitta/gpu-unlock}}<br />{{Source\|handle=zagitta\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563543666553913345}}<br />{{Source\|handle=zagitta\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563559641152098425}}<br />{{Source\|handle=zagitta\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563585750836969533}}<br />{{Source\|handle=zagitta\|date=2026-10-25\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563875998443765832}}
			** {{Status\|Reported}} @danske1's before and after numbers on an EVGA RTX 3090 FTW3, 8192 x 8192 x 8192 GEMM, median of 20 iterations x 5 samples: FP16 in with FP32 accumulate 81.95 to 141.07 TFLOPS (+72 %), BF16 / FP32 81.35 to 145.59 (+79 %), TF32 40.46 to 78.76 (+95 %); FP16 with FP16 accumulate (about 148) and plain FP32 (about 30) unchanged. {{Source\|handle=danske1\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563293585762295871}}<br />{{Source\|handle=danske1\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563293742557962313}}
			* {{Status\|Reported}} '''JustAResearcher/gpu-unlock-ada''' ([https://github.com/JustAResearcher/gpu-unlock-ada GitHub]) takes the tensor unlock to RTX 40-series (Ada) cards. @petrikrohn_26497 reported an RTX 4090 unlocked and posted the repo. Its README credits Zagitta/gpu-unlock for the Linux module and packaging, TOTORONG/AD103_win_unlocker for the AD102 support and WildFlash1st/cmp90hx-unlock for the GA102 path; the module and packaging are GPL-2.0-only and the NVIDIA firmware stays under NVIDIA's own terms. @djpandaman, earlier the same day: the tensor unlock can be done on AD103 too, and he was trying AD102. @sevendaysworth unlocked an RTX 4080 FE by pointing an AI agent at another member's repo (the repo is not named in his post). Not a CMP 170HX tool. A day earlier @zagitta had posted TOTORONG/AD103_win_unlocker on Hugging Face ("Seems someone got ada unlock working maybe") and @dottoreargus replied "Can confirm, it does work". {{Source\|handle=zagitta\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563248053916536933}}<br />{{Source\|handle=dottoreargus\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563314008717791445}}<br />{{Source\|handle=petrikrohn_26497\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563577439504830544}}<br />{{Source\|handle=petrikrohn_26497\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563578905913331836\|repo=https://github.com/JustAResearcher/gpu-unlock-ada}}<br />{{Source\|handle=djpandaman\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563559770307436645}}<br />{{Source\|handle=sevendaysworth\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563612387800522785}}
			* Open question: has anyone tried the unlock on a GA104 laptop GPU? @da6b asked on 2026-10-23; no answer in the thread. {{Source\|handle=da6b\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563174949785698396}}
	* {{Status\|Theory}} @djpandaman asked AD102 owners (4090, RTX 5000 Ada, 6000 Ada) to run a read-only fuse dump script; @zagitta posted fuse and floorsweep register values from two RTX 3090s for reference. {{Source\|handle=zagitta\|date=2026-10-26\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1564323744510648351}}<br />{{Source\|handle=djpandaman\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1564512498743378011}}<br />{{Source\|handle=djpandaman\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1564512653253148692}}		* {{Status\|Theory}} @djpandaman asked AD102 owners (4090, RTX 5000 Ada, 6000 Ada) to run a read-only fuse dump script; @zagitta posted fuse and floorsweep register values from two RTX 3090s for reference. {{Source\|handle=zagitta\|date=2026-10-26\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1564323744510648351}}<br />{{Source\|handle=djpandaman\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1564512498743378011}}<br />{{Source\|handle=djpandaman\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1564512653253148692}}
	* Reverse-engineering links people shared: a laser fault injection rig build (fraktalcyber/lfi-rig) and an agent tool whose list lacks NVIDIA Falcon support; an agent with Ghidra and radare2 can disassemble most things. {{Source\|handle=mountains1605\|date=2026-10-09\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1558202438660657193}}<br />{{Source\|handle=analyticeth\|date=2026-10-09\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1558190561977045017}}<br />{{Source\|handle=dottoreargus\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1559852208168771646}}		* Reverse-engineering links people shared: a laser fault injection rig build (fraktalcyber/lfi-rig) and an agent tool whose list lacks NVIDIA Falcon support; an agent with Ghidra and radare2 can disassemble most things. {{Source\|handle=mountains1605\|date=2026-10-09\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1558202438660657193}}<br />{{Source\|handle=analyticeth\|date=2026-10-09\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1558190561977045017}}<br />{{Source\|handle=dottoreargus\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1559852208168771646}}

ImportBot: Import: tools overview, @mountains1605, @globalapathy, @xeax, @matstatman, @c133, @wakeupmbrc1, @zagitta, @impactic, @djpandaman, @sojun80, @arabus_74064, @reala7096, @_arcanum_, @rh3dai; repos PixelML/club-170hx, SeraphimSerapis/tool-eval-bench, dev-boz/cmp170hx-testsuite, lixiangwuxian/170hx-rir-tool, sasha0552/nvidia-pstated, zzc0721/torch-performance-test-data

2026-10-30T08:53:04Z

@@ Line 113: / Line 113: @@
 * Nobody has a fork that combines P2P and [[170tune]]. {{Source|handle=gnasherrr|date=2026-10-26|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1564165428631179337}}
-== Longer write-ups on Discord ==
+== Write-ups on Discord ==
-These are summarised here only until their authors agree to more.
+Summaries of longer write-ups members posted, with links to the originals.
-* <code><nowiki>iff_apply.py</nowiki></code>, <code><nowiki>170hx_sweep.py</nowiki></code> and <code><nowiki>bar0_write.py</nowiki></code> with their READMEs, posted as "untested": a longer write-up by @xeax and @scallion_9883 (Humvee55) exists; see [https://discord.com/channels/1518634379604136178/1518692084729511967/1562866846083055622 Discord].
-* Building cmpunlocker's vGPU-and-others branch: a longer write-up by @arabus_74064 (arabus) exists; see [https://discord.com/channels/1518634379604136178/1518692084729511967/1564633109482438887 Discord].
+=== 170hx_sweep.py, bar0_write.py and iff_apply.py (10 GB card) ===
-* Speeding up model loading at startup: a longer write-up by @impactic (Impactic) exists; see [https://discord.com/channels/1518634379604136178/1528826330454429706/1560848620633198652 Discord].
+{{Status|Theory}} @xeax posted <code><nowiki>iff_apply_README.md</nowiki></code>, <code><nowiki>170hx_sweep_README.md</nowiki></code>, <code><nowiki>bar0_write.py</nowiki></code> and <code><nowiki>170hx_sweep.py</nowiki></code> with one word: "untested". {{Source|handle=xeax|date=2026-10-22|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1562866846083055622}}<br />{{Source|handle=xeax|date=2026-10-22|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1562869594794754209}}<br />{{Source|handle=xeax|date=2026-10-22|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1562869675640229989}}<br />{{Source|handle=xeax|date=2026-10-22|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1562870077097910425}} @scallion_9883 (Humvee55) had shared a <code><nowiki>message.txt</nowiki></code> when @feralzi asked for <code><nowiki>iff_apply.py</nowiki></code>, and confirmed the tools are "10gb only". {{Source|handle=scallion_9883|date=2026-10-22|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1562853684822413492}}<br />{{Source|handle=scallion_9883|date=2026-10-22|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1562937468754329762}} What the sweep README describes:
-* vLLM checkpointing and fast model swapping: a longer write-up by @peter643 exists; see [https://discord.com/channels/1518634379604136178/1528513735575081070/1557509343094509569 Discord].
+* '''Purpose.''' It tests up to 128 SMs and up to 24 HBM memory partitions, including the ones fused off at the factory, one unbind, reset, override, bind and test cycle at a time, and writes <code><nowiki>CARD_REPORT.md</nowiki></code>: a verdict per GPC, TPC and memory partition, the defective units, and the masks for the largest stable configuration of that die. Every test uses volatile overrides through the <code><nowiki>ga100_rir</nowiki></code> binary of [[170hx-rir-tool]], so a reboot or power cycle returns the card to stock; the README says no RIR write path exists in the code. {{Source|handle=xeax|date=2026-10-22|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1562869594794754209}} The tool's software overrides only work once the gate fuse is cleared; see [[170hx-rir-tool]]. {{Source|handle=lxtend|date=2026-10-10|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1558470807888461854|repo=https://github.com/lixiangwuxian/170hx-rir-tool}}.
-* A video series on tuning the 8 GB card and on cmpunlocker v0.5: a longer write-up by @rh3dai (RH-3D) exists; see [https://discord.com/channels/1518634379604136178/1555480078089986138/1556301132861800551 Discord].
+* '''Requirements.''' Debian or Ubuntu, root, one card per run, no display session on the card, a working NVIDIA driver. gpu-burn, cuda_memtest and torch are needed for real verdicts; without them tests are logged as skipped. Memory stages above 40 GiB need the patched cmpunlocker driver with the live-FBPA LMR fix (<code><nowiki>--assume-memory-driver</nowiki></code>); otherwise they are skipped. By default it only accepts device IDs 0x2082 and 0x2081. {{Source|handle=xeax|date=2026-10-22|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1562869594794754209}}
-* Driving a PC with an AI agent through an IP-KVM: a longer write-up by @con3po7016 (Con3po) exists; see [https://discord.com/channels/1518634379604136178/1518691940646781070/1558269252111306807 Discord].
+* '''Commands.''' <code><nowiki>sudo python3 170hx_sweep.py discover</nowiki></code> (read-only scan of fuses and masks), <code><nowiki>run</nowiki></code> (resumable; about 35 short GPU cycles, 45 to 60 minutes), <code><nowiki>status</nowiki></code>, <code><nowiki>report</nowiki></code>. Useful flags: <code><nowiki>--quick</nowiki></code> (skip the per-TPC matrix), <code><nowiki>--skip-memory</nowiki></code>, <code><nowiki>--verbose</nowiki></code>. Watch <code><nowiki>dmesg -w</nowiki></code> and <code><nowiki>nvidia-smi -l 1</nowiki></code> in another terminal. {{Source|handle=xeax|date=2026-10-22|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1562869594794754209}}
 == See also ==

ImportBot: Import: tools overview, @mountains1605, @globalapathy, @c133, @wakeupmbrc1, @zagitta, @djpandaman, @sojun80, @matstatman, @reala7096, @_arcanum_, @xanthefin, @nestaa51, @boz._., @xeax; repos PixelML/club-170hx, SeraphimSerapis/tool-eval-bench, dev-boz/cmp170hx-testsuite, sasha0552/nvidia-pstated, zzc0721/torch-performance-test-data

2026-10-30T08:05:12Z

Import: tools overview, @mountains1605, @globalapathy, @c133, @wakeupmbrc1, @zagitta, @djpandaman, @sojun80, @matstatman, @reala7096, @_arcanum_, @xanthefin, @nestaa51, @boz._., @xeax; repos PixelML/club-170hx, SeraphimSerapis/tool-eval-bench, dev-boz/cmp170hx-testsuite, sasha0552/nvidia-pstated, zzc0721/torch-performance-test-data

Show changes

ImportBot: Wiki lane: site structure (templates, forms, help pages)

2026-10-30T07:00:57Z

Wiki lane: site structure (templates, forms, help pages)

New page

''This section is being written from the Discord history.'' It will cover monitoring, benchmarks and scripts.

Until then, see [[Start here]] and [[Contributing]].

← Older revision		Revision as of 10:02, 30 October 2026
Line 85:		Line 85:
	* {{Status\|Reported}} '''vLLM run-to-run variance.''' Re-running the same quantization comparison gave results about 3 percent apart; @vektor_prime traced it to upstream vLLM and fixed it with a batch-invariance change in his recipe repository. {{Source\|handle=vektor_prime\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1557467960455467170}}<br />{{Source\|handle=vektor_prime\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1557498798530498601}}		* {{Status\|Reported}} '''vLLM run-to-run variance.''' Re-running the same quantization comparison gave results about 3 percent apart; @vektor_prime traced it to upstream vLLM and fixed it with a batch-invariance change in his recipe repository. {{Source\|handle=vektor_prime\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1557467960455467170}}<br />{{Source\|handle=vektor_prime\|date=2026-10-07\|link=https://discord.com/channels/1518634379604136178/1518695289278566450/1557498798530498601}}
	* {{Status\|Contested}} '''LMCache''' for spilling KV cache to CPU RAM: with it vLLM processed one request at a time instead of eight, and @c133 found it also wanted VRAM. {{Source\|handle=nikhilr1696\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1556742092150284339}}<br />{{Source\|handle=c133\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1556741812394270760}}		* {{Status\|Contested}} '''LMCache''' for spilling KV cache to CPU RAM: with it vLLM processed one request at a time instead of eight, and @c133 found it also wanted VRAM. {{Source\|handle=nikhilr1696\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1556742092150284339}}<br />{{Source\|handle=c133\|date=2026-10-05\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1556741812394270760}}
			* {{Status\|Reported}} '''NeuralFree''' ([https://github.com/EonMux/NeuralFree EonMux/NeuralFree]): a work-in-progress inference framework by @phoenixone7. GGUF support is in; safetensors support and testing are still to come. He split it out of an older trading-bot project and plans to keep supporting it because of prefix-caching problems he hit in vLLM. No CMP numbers posted. {{Source\|handle=phoenixone7\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563213261850157077\|repo=https://github.com/EonMux/NeuralFree}}<br />{{Source\|handle=phoenixone7\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563213796913188956}}
	* {{Status\|Theory}} '''EAGLE-4''' speculative decoding was flagged as faster than MTP; nobody tried it on a 170HX. {{Source\|handle=cibernox\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562247115562942555}}		* {{Status\|Theory}} '''EAGLE-4''' speculative decoding was flagged as faster than MTP; nobody tried it on a 170HX. {{Source\|handle=cibernox\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562247115562942555}}
	* {{Status\|Theory}} '''flare-local-inference''' (NVIDIA) is built for Blackwell; a port to these cards would need a purpose-written engine. {{Source\|handle=mountains1605\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562492007262322789}}<br />{{Source\|handle=dufrelic\|date=2026-10-22\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562824922286526505}}		* {{Status\|Theory}} '''flare-local-inference''' (NVIDIA) is built for Blackwell; a port to these cards would need a purpose-written engine. {{Source\|handle=mountains1605\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562492007262322789}}<br />{{Source\|handle=dufrelic\|date=2026-10-22\|link=https://discord.com/channels/1518634379604136178/1528826330454429706/1562824922286526505}}
Line 96:		Line 97:
	* Iron: a Pinecil or any 60 W or stronger iron with SMD tips; the Sequre S99 (USB-C, C245 tips, 100 W). {{Source\|handle=civiloid\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558803228362145833}}<br />{{Source\|handle=globalapathy\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558832357514154055}}		* Iron: a Pinecil or any 60 W or stronger iron with SMD tips; the Sequre S99 (USB-C, C245 tips, 100 W). {{Source\|handle=civiloid\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558803228362145833}}<br />{{Source\|handle=globalapathy\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558832357514154055}}
	* Microsoldering tweezers are essential; the capacitors are almost impossible to handle otherwise. Clean with isopropyl alcohol, foam swabs, a brush and wipes. {{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558857956961820794}}<br />{{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558858884343668859}}		* Microsoldering tweezers are essential; the capacitors are almost impossible to handle otherwise. Clean with isopropyl alcohol, foam swabs, a brush and wipes. {{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558857956961820794}}<br />{{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558858884343668859}}
	* Microscope: @civiloid would rather have a QianLi Gorilla 6558 or Mechanic 6560T with a camera than work off a screen alone. {{Source\|handle=civiloid\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558803596852596837}} A used Vision Engineering TS-4 is another option; skip the TS-3. {{Source\|handle=cmdrwaka324\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559782730688893020}}		* Microscope: @civiloid would rather have a QianLi Gorilla 6558 or Mechanic Robot 6560T with a camera than work off a screen alone. {{Source\|handle=civiloid\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558803596852596837}} For soldering he calls a new Chinese binocular microscope with a camera module the most cost-effective choice: small, good enough optics, and the camera plugs into any screen. QianLi and Mechanic are consumer brands that buy white-label units, so the actual manufacturer is cheaper. {{Source\|handle=civiloid\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559808745176043613}}<br />{{Source\|handle=civiloid\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559808927581872161}}<br />{{Source\|handle=civiloid\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559809944977408102}} @_arcanum_ paid about USD 300 on Taobao for the QianLi 6558 UV-light version; the IR module is about 140 more. {{Source\|handle=_arcanum_\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559983772076609640}}<br />{{Source\|handle=_arcanum_\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559985698352795670}}<br />{{Source\|handle=_arcanum_\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559985805802340454}} A used Vision Engineering TS-4 is another option; skip the TS-3, which is louder (fans on the supply and the halogen housing) and older in optics and electronics. {{Source\|handle=cmdrwaka324\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559782730688893020}}<br />{{Source\|handle=cmdrwaka324\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559798125542248489}} On the TS series, video output needs the camera optics (the CS adapter) fitted when bought; scopes in this class run USD 2,000 to 3,000, and eyepiece-only scopes cost about 5x less. Many cheap camera heads send one eye's light fully to the camera, so you lose binocular view while recording. {{Source\|handle=endsjeremy\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559786404223914037}}<br />{{Source\|handle=cmdrwaka324\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559788016095076363}}<br />{{Source\|handle=endsjeremy\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559804501127462943}}<br />{{Source\|handle=endsjeremy\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559805792067125259}}<br />{{Source\|handle=endsjeremy\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1559984711525343233}}
	* A solder stencil for the mod is in the channel pins, but making one needs a laser cutter. {{Source\|handle=globalapathy\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1564479737160802307}}<br />{{Source\|handle=globalapathy\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1564479895952695377}}		* A solder stencil for the mod is in the channel pins, but making one needs a laser cutter. {{Source\|handle=globalapathy\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1564479737160802307}}<br />{{Source\|handle=globalapathy\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1564479895952695377}}
	* A video of the mod done quickly by a Canadian GPU repair YouTuber: https://www.youtube.com/watch?v=HO2T4TgNzlQ {{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558856094116483142}}		* A video of the mod done quickly by a Canadian GPU repair YouTuber: https://www.youtube.com/watch?v=HO2T4TgNzlQ {{Source\|handle=_arcanum_\|date=2026-10-11\|link=https://discord.com/channels/1518634379604136178/1534370835220004885/1558856094116483142}}
Line 109:		Line 110:
	== Related work on other NVIDIA cards ==		== Related work on other NVIDIA cards ==
	* {{Status\|Reported}} '''Zagitta/gpu-unlock''': a Linux kernel module that removes the half-rate tensor-math limit on RTX 3090 class cards with the register approach the CMP unlocks use, without driver patching. It must load before the NVIDIA driver; it only matches RTX 3090 IDs by default. Gain: a few percent prefill when the engine uses BF16 to FP32 FMA; no core or SM unlock. {{Source\|handle=zagitta\|date=2026-10-18\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1561189774834802709}}<br />{{Source\|handle=zagitta\|date=2026-10-19\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1561876551505551411}}<br />{{Source\|handle=zagitta\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562214717316137005}}<br />{{Source\|handle=zagitta\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563321762756694107}} @sojun80 used the approach on a 3060 Ti, 3070, 3070 mobile, 3080 and 3080 Ti, and is building a generic unlock tool from cmpunlocker for the 30 and 40 series. {{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562322577676967957}}<br />{{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562322921383264289}} Details belong on [[Unlocks]].		* {{Status\|Reported}} '''Zagitta/gpu-unlock''': a Linux kernel module that removes the half-rate tensor-math limit on RTX 3090 class cards with the register approach the CMP unlocks use, without driver patching. It must load before the NVIDIA driver; it only matches RTX 3090 IDs by default. Gain: a few percent prefill when the engine uses BF16 to FP32 FMA; no core or SM unlock. {{Source\|handle=zagitta\|date=2026-10-18\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1561189774834802709}}<br />{{Source\|handle=zagitta\|date=2026-10-19\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1561876551505551411}}<br />{{Source\|handle=zagitta\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562214717316137005}}<br />{{Source\|handle=zagitta\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563321762756694107}} @sojun80 used the approach on a 3060 Ti, 3070, 3070 mobile, 3080 and 3080 Ti, and is building a generic unlock tool from cmpunlocker for the 30 and 40 series. {{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562322577676967957}}<br />{{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562322921383264289}} Details belong on [[Unlocks]].
			** @grapepepper patched his RTX 3090 Ti from the repository; after he shared his results others unlocked various RTX 30-series cards, and he has seen a report of an RTX 4070 Ti SUPER. {{Source\|handle=grapepepper\|date=2026-10-20\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562183666371924070}}
			** '''eGPU.''' Zagitta: someone got it working on an eGPU (see the repo issues). {{Status\|Theory}} @da6b, whose 3090 sits behind a Thunderbolt 3 bridge, expects to need the Thunderbolt drivers in the initramfs and Thunderbolt security disabled so the tunnel is up before the module loads, possibly with some timing involved; an eGPU on M.2 is a direct PCIe link and simpler. {{Source\|handle=zagitta\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563321076027752459}}<br />{{Source\|handle=da6b\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563320911858376785}}<br />{{Source\|handle=da6b\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563321227320492143}}
			** {{Status\|Theory}} @sojun80 thinks the tensor limit applies to the 20 to 40 series, "i think all of them"; he has no 20-series card left to test. {{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562275001879629885}}<br />{{Source\|handle=sojun80\|date=2026-10-21\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1562322138516684843}}
			** Zagitta on its origin: it builds on the CMP unlock work (the same basis as cmpunlocker, and the CMP 90HX and 50HX unlocks), and nothing in it is new; his only change is doing it without a driver patch, as a standalone early-boot module, an idea he took from the UEFI module the 90HX and 50HX Windows unlocks use. He has no Windows version planned. On 2026-10-24, asked whether it covers only GA102, he said GA104 and GA106 as well; other GPUs in the same host (a 50-series card, for example) are left alone. {{Source\|handle=zagitta\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563146783062884393}}<br />{{Source\|handle=zagitta\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563529939586912410}}<br />{{Source\|handle=mrwojaketh\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563529790848372796\|repo=https://github.com/Zagitta/gpu-unlock}}<br />{{Source\|handle=zagitta\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563543666553913345}}<br />{{Source\|handle=zagitta\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563559641152098425}}<br />{{Source\|handle=zagitta\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563585750836969533}}<br />{{Source\|handle=zagitta\|date=2026-10-25\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563875998443765832}}
			** {{Status\|Reported}} @danske1's before and after numbers on an EVGA RTX 3090 FTW3, 8192 x 8192 x 8192 GEMM, median of 20 iterations x 5 samples: FP16 in with FP32 accumulate 81.95 to 141.07 TFLOPS (+72 %), BF16 / FP32 81.35 to 145.59 (+79 %), TF32 40.46 to 78.76 (+95 %); FP16 with FP16 accumulate (about 148) and plain FP32 (about 30) unchanged. {{Source\|handle=danske1\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563293585762295871}}<br />{{Source\|handle=danske1\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563293742557962313}}
			* {{Status\|Reported}} '''JustAResearcher/gpu-unlock-ada''' ([https://github.com/JustAResearcher/gpu-unlock-ada GitHub]) takes the tensor unlock to RTX 40-series (Ada) cards. @petrikrohn_26497 reported an RTX 4090 unlocked and posted the repo. Its README credits Zagitta/gpu-unlock for the Linux module and packaging, TOTORONG/AD103_win_unlocker for the AD102 support and WildFlash1st/cmp90hx-unlock for the GA102 path; the module and packaging are GPL-2.0-only and the NVIDIA firmware stays under NVIDIA's own terms. @djpandaman, earlier the same day: the tensor unlock can be done on AD103 too, and he was trying AD102. @sevendaysworth unlocked an RTX 4080 FE by pointing an AI agent at another member's repo (the repo is not named in his post). Not a CMP 170HX tool. A day earlier @zagitta had posted TOTORONG/AD103_win_unlocker on Hugging Face ("Seems someone got ada unlock working maybe") and @dottoreargus replied "Can confirm, it does work". {{Source\|handle=zagitta\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563248053916536933}}<br />{{Source\|handle=dottoreargus\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1563314008717791445}}<br />{{Source\|handle=petrikrohn_26497\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563577439504830544}}<br />{{Source\|handle=petrikrohn_26497\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563578905913331836\|repo=https://github.com/JustAResearcher/gpu-unlock-ada}}<br />{{Source\|handle=djpandaman\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563559770307436645}}<br />{{Source\|handle=sevendaysworth\|date=2026-10-24\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563612387800522785}}
			* Open question: has anyone tried the unlock on a GA104 laptop GPU? @da6b asked on 2026-10-23; no answer in the thread. {{Source\|handle=da6b\|date=2026-10-23\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1563174949785698396}}
	* {{Status\|Theory}} @djpandaman asked AD102 owners (4090, RTX 5000 Ada, 6000 Ada) to run a read-only fuse dump script; @zagitta posted fuse and floorsweep register values from two RTX 3090s for reference. {{Source\|handle=zagitta\|date=2026-10-26\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1564323744510648351}}<br />{{Source\|handle=djpandaman\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1564512498743378011}}<br />{{Source\|handle=djpandaman\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1564512653253148692}}		* {{Status\|Theory}} @djpandaman asked AD102 owners (4090, RTX 5000 Ada, 6000 Ada) to run a read-only fuse dump script; @zagitta posted fuse and floorsweep register values from two RTX 3090s for reference. {{Source\|handle=zagitta\|date=2026-10-26\|link=https://discord.com/channels/1518634379604136178/1518692084729511967/1564323744510648351}}<br />{{Source\|handle=djpandaman\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1564512498743378011}}<br />{{Source\|handle=djpandaman\|date=2026-10-27\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1564512653253148692}}
	* Reverse-engineering links people shared: a laser fault injection rig build (fraktalcyber/lfi-rig) and an agent tool whose list lacks NVIDIA Falcon support; an agent with Ghidra and radare2 can disassemble most things. {{Source\|handle=mountains1605\|date=2026-10-09\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1558202438660657193}}<br />{{Source\|handle=analyticeth\|date=2026-10-09\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1558190561977045017}}<br />{{Source\|handle=dottoreargus\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1559852208168771646}}		* Reverse-engineering links people shared: a laser fault injection rig build (fraktalcyber/lfi-rig) and an agent tool whose list lacks NVIDIA Falcon support; an agent with Ghidra and radare2 can disassemble most things. {{Source\|handle=mountains1605\|date=2026-10-09\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1558202438660657193}}<br />{{Source\|handle=analyticeth\|date=2026-10-09\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1558190561977045017}}<br />{{Source\|handle=dottoreargus\|date=2026-10-14\|link=https://discord.com/channels/1518634379604136178/1518695328008638516/1559852208168771646}}