[ANN] FreeForGemVela 0.4.6 - VELA GPU miner | 2 lanes | AMD RX 6800 XT | CMP 50HX | auto-undervolt |

[pepsykolya]

FreeForGemVela (ffgv) 0.4.6
GPU miner for VELA (Proof-of-Inference, integer sd-turbo)
NVIDIA RTX 30 / 40 / 50 · CMP 50HX · AMD RX 6800 XT · Windows · HiveOS · Linux · dev fee 1%

0.4.6 (12.11.2026): two lanes on the RTX 3070 (desktop and laptop), RTX 4070 Ti and CMP 50HX, batch 20 on the RTX 3080 Ti. Two lanes are two inference streams on one GPU; where it was measured to help, the lanes start every batch together (launch barrier). Measured on the miner's own inf/s against 0.4.5: RTX 3070 +6.1..6.3%, RTX 3070 Laptop +4.8..5.1% (34.1 inf/s), RTX 3080 Ti +10.5% (batch 20, one lane, 64.9 inf/s), RTX 4070 Ti +2.4% (73.9 inf/s with 2 x batch 6, joules per inference about 2% worse), CMP 50HX +4% (a rig of 5 cards: 159.9 inf/s). A lane that hangs or fails its start check is stopped alone and the GPU goes on with the others. --lanes N sets the number by hand, --no-lanes (= --lanes 1) switches it off. The math of the NVIDIA kernels is that of 0.4.5 (bit-exact, 56 canaries on every lane); the AMD engine is unchanged. Release notes: https://github.com/pepsykolya/FreeForGemVela/releases/tag/v0.4.6

0.4.5 (07.11.2026): fixes. Solo mining on your own node reacts to a new block faster: submitblock no longer stops the polling for the next block (a solo miner saw new jobs 15-17 s late), a node of your own (local network, Tailscale) is polled every 0.5 s, the RPC connection is kept (keep-alive); measured on a rig on another site: a new job within 0-1 s of a new block instead of 4-5 s. HiveOS: "need update!!" next to the version when a newer release is out. suprnova is the first pool in the examples. The engines (NVIDIA and AMD) are unchanged from 0.4.4. The sending of a found block is so far only unit-tested, not on a real node.
0.4.4 (05.11.2026): the miner downloads its model itself, and the first AMD Radeon support: RX 6800 / 6800 XT / 6900 XT / 6950 XT (Navi 21), Windows. The NVIDIA engine code is unchanged (as in 0.4.3). AMD: OpenCL engine, nothing to install (no HIP / ROCm), bit-exact with the node consensus, checked at every start; the miner sets clock, voltage and power limit and always puts the owner's values back. RX 6800 XT, batch 16: 23.2 inf/s at 272 W (stock), 22.0 inf/s at 198 W in --mode eco (-27% power, +30% inf/s per W). Undervolting a Radeon is risky: past its edge it computes wrong results silently, then the video driver hangs - the presets stay on the safe side and every step is checked by canaries (read the manual). AMD is Windows only in this release; the Linux / HiveOS packages have no AMD engine yet and other Radeon chips are not supported yet. If the model file is missing at the start, ffgv downloads it from Hugging Face (mirror as a fallback, resumes a broken download, sha256 checked; --no-model-download turns it off). CMP 50HX (0.4.3): on HiveOS a CMP card needs a driver that unlocks it (without it about 11 times slower): --cmp-unlock-driver in Extra config or sudo ./ffgv --cmp-driver; it is a third-party driver build, replaces the NVIDIA driver of the whole rig and reboots it - read the manual (section 6) first.
VirusTotal, 0.4.6 packages: Windows https://www.virustotal.com/gui/file/59ac276982c732644442b3c8632af06bccaf65d6f1bfb7f8f8a39501906dfccc (4 detections: huorong HackTool/CoinMiner, Kaspersky not-a-virus RiskTool.BAT, Sangfor and DeepInstinct generic heuristics; typical for a miner with .bat start scripts; 63 engines clean, 7 do not support the file type, 1 timed out) · HiveOS https://www.virustotal.com/gui/file/57814f5e25300542fa5c72d260b81a950c669498f12fb2ed39becae72583c6ca (0 detections; 62 engines clean, 13 do not support the file type) · Linux https://www.virustotal.com/gui/file/5ac9a6e17551a57f923f56730e7b33c0fa618f45d984daf1fca86f6253d3bcb2 (0 detections; 62 engines clean, 13 do not support the file type)

Download: https://github.com/pepsykolya/FreeForGemVela/releases/tag/v0.4.6
Manual (EN/RU): https://github.com/pepsykolya/FreeForGemVela
Video tutorial, 0.4.6 (3 parts, 10-14 min each):
EN: Part 1 (the coin) https://youtu.be/WaemJ-atD3I · Part 2 (the miner: start, extra arguments) https://youtu.be/OHWl6AFum2Y · Part 3 (lanes, undervolt) https://youtu.be/N7Gp00v4JE8
RU: Part 1 https://youtu.be/bSCrNJ2I9UE · Part 2 https://youtu.be/KVPm6pZtkrM · Part 3 https://youtu.be/iplVsjnC_OQ

What it does

• Every inference is computed bit-exact to the VELA node's consensus code; each share is re-checked before it is sent.
• Auto-undervolt that is safe on VELA: an undervolted GPU here does not crash first - it silently computes a wrong result and the pool rejects the share. ffgv tests every step byte for byte against known consensus results (56 canaries), keeps checking while mining and steps back on any mismatch.
• Ready modes --mode speed | balance | eco; hang watchdog; Xid self-recovery; pool failover; stratum over TLS; --suggest-diff; JSON API.
• HiveOS custom miner (miner name ffgv-hive), Windows start.bat (downloads and checks the model itself), Linux.
• CMP 50HX (Turing): own --mode presets and auto-undervolt start; on HiveOS the CMP unlock driver can be installed on request (--cmp-unlock-driver).
• Two lanes per GPU where they pay off (RTX 3070 desktop and laptop, RTX 4070 Ti, CMP 50HX) and batch 20 on the RTX 3080 Ti: +2..+10% inf/s against 0.4.5; a hung or failing lane is stopped alone. --lanes N / --no-lanes.

Speed and power (RTX rows: release 0.4.2, mining on AriaPool; CMP and AMD rows: releases 0.4.3 / 0.4.4, offline benchmark at batch 16 with the CMP unlock driver; the power limit is only a safety cap, the mode sets the clocks)

Code:

GPU (power limit)          mode            inf/s    W      inf/s/W
RTX 3070 (200 W cap)       balance + UV    36.02   122.3   0.295
RTX 3070 (200 W cap)       eco + UV        32.71   100.0   0.327
RTX 4070 Ti (235 W)        stock           72.82   234.1   0.311
RTX 4070 Ti (235 W)        speed + UV      75.09   234.2   0.321
RTX 4070 Ti (235 W)        balance         66.17   178.5   0.371
RTX 4070 Ti (235 W)        eco             54.01   141.1   0.383
RTX 3080 Ti (350 W), 0.4.0 stock           61.45   334.6   0.184
RTX 3080 Ti (350 W), 0.4.0 eco + UV        49.35   209.6   0.235
4x RTX 3070, HiveOS        balance + UV   142.1    464     0.307
CMP 50HX (225 W), 0.4.3    stock           34.46   219.0   0.157
CMP 50HX (225 W), 0.4.3    1650 lock, +240 35.10   182.9   0.192
CMP 50HX (225 W), 0.4.3    balance + UV    32.61   150.8   0.216
CMP 50HX (225 W), 0.4.3    eco + UV        29.64   134.5   0.220
RX 6800 XT (Windows), 0.4.4   stock 2324 MHz/1150 mV  23.22   271.8   0.085
RX 6800 XT (Windows), 0.4.4   2324 MHz/1050 mV   23.38   230.2   0.102
RX 6800 XT (Windows), 0.4.4   eco 2150 MHz/881 mV 21.98   198.1   0.111

Full table and method: https://github.com/pepsykolya/FreeForGemVela/blob/main/docs/SPEED.md
The table is of 0.4.2-0.4.4 modes; the 0.4.6 two-lane / batch-20 figures are in the first paragraph above.

Pools (tested, 0 rejected)

Code:

AriaPool   http://159.69.194.46:3349
suprnova   stratum+tcp://vela.suprnova.cc:7601   (vardiff)
           stratum+tcp://vela.suprnova.cc:7602   (fixed 2048, one GPU)
           stratum+tcp://vela.suprnova.cc:7603   (fixed 16384, rigs)
           stratum+ssl://vela.suprnova.cc:7604   (vardiff over TLS)
solo       http://RPCUSER:RPCPASS@127.0.0.1:19732 (your own VELA node)

Login on the pools: WALLET.WORKER, no registration. Several --pool = failover.

Quick start
HiveOS: Custom miner -> Installation URL https://github.com/pepsykolya/FreeForGemVela/releases/download/v0.4.6/ffgv-hive-0.4.6.tar.gz , miner name ffgv-hive, wallet %WAL%.%WORKER_NAME%, pool URL from the list above, extra config e.g. --mode balance --pl 200 (two lanes start by themselves on an RTX 3070, RTX 4070 Ti and CMP 50HX, --no-lanes switches them off) (speed/balance lift a HiveOS memory lock by themselves).
Windows: unpack, right-click start.bat -> Edit -> WALLET / POOL / WORKER / MODE, run as administrator.
Linux: ./ffgv --pool stratum+ssl://vela.suprnova.cc:7604 --user vela1...YOUR_WALLET.rig1 --model sd-turbo-int.poisd --mode balance

VELA: https://velaresearch.ai · Discord https://discord.gg/VcTU8h3pv7 · X https://x.com/velaresearch_ai
Our Telegram: https://t.me/openclaw_ru_train (chat https://t.me/OpenClaw_RUTrain)

Dev fee 1% (one 60 s slot every 100 minutes). License: free to use and redistribute unmodified.

Pyccкий

FreeForGemVela (ffgv) 0.4.6 - GPU-мaйнep для VELA (Proof-of-Inference). NVIDIA RTX 30 / 40 / 50, AMD RX 6800 XT (Windows), Windows, HiveOS, Linux, кoмиccия 1%.

• Кaждaя инфepeнция пoбaйтнo coвпaдaeт c кoнceнcycoм нoды, кaждaя шapa пepeпpoвepяeтcя пepeд oтпpaвкoй.
• Бeзoпacный aвтoaндepвoльт: нa VELA кapтa пpи cлишкoм глyбoкoм aндepвoльтe cнaчaлa нe пaдaeт, a тиxo cчитaeт нeвepный peзyльтaт, и пyл oтклoняeт шapy. ffgv пpoвepяeт кaждyю cтyпeнь пoбaйтнo пo 56 этaлoнным инфepeнциям, пpoдoлжaeт пpoвepять вo вpeмя мaйнингa и oтcтyпaeт пpи любoм pacxoждeнии.
• Peжимы --mode speed | balance | eco, cтopoж зaвиcaний, caмoвoccтaнoвлeниe пocлe Xid, peзepвныe пyлы, stratum пo TLS, JSON API.
• Hoвoe в 0.4.4: мaйнep caм cкaчивaeт мoдeль c Hugging Face (c зepкaлoм, дoкaчкa, пpoвepкa sha256), ecли фaйлa нeт (--no-model-download oтключaeт). Пepвaя пoддepжкa AMD Radeon RX 6800 / 6800 XT / 6900 XT / 6950 XT (Navi 21), тoлькo Windows: движoк нa OpenCL, ничeгo cтaвить нe нyжнo, peзyльтaты пoбaйтнo кaк y кoнceнcyca. RX 6800 XT: 23,2 inf/s пpи 272 Bт (штaтнo), 22,0 inf/s пpи 198 Bт в --mode eco. Aндepвoльт Radeon pиcкoвaн (пpи пepeбepe кapтa тиxo cчитaeт нeвepнo, зaтeм виcнeт дpaйвep) - пpeceты кoнcepвaтивныe, читaйтe pyкoвoдcтвo. Пaкeты Linux / HiveOS пoкa бeз AMD.
• CMP 50HX (Turing): cвoи пpeceты --mode; нa HiveOS пo зaпpocy мoжнo пocтaвить дpaйвep-paзблoкиpoвкy CMP (--cmp-unlock-driver), бeз нeгo кapтa CMP дaёт пpимepнo в 11 paз мeньшe.
• 0.4.6: двe лeйны нa RTX 3070 (desktop и laptop), RTX 4070 Ti и CMP 50HX, batch 20 нa RTX 3080 Ti: +2..+10% inf/s пpoтив 0.4.5 (измepeнo пo inf/s мaйнepa). --lanes N - выбop вpyчнyю, --no-lanes - выключить.

Cкaчaть: https://github.com/pepsykolya/FreeForGemVela/releases/tag/v0.4.6
Maнyaл (RU): https://github.com/pepsykolya/FreeForGemVela/blob/main/MANUAL.ru.md
Bидeoypoк RU (0.4.6, 3 чacти): https://youtu.be/bSCrNJ2I9UE · https://youtu.be/KVPm6pZtkrM · https://youtu.be/iplVsjnC_OQ
Taблицa cкopocти и пyлы - вышe. Лoгин нa пyлax: КOШEЛЁК.BOPКEP, бeз peгиcтpaции.
Haш Telegram: https://t.me/openclaw_ru_train

[pepsykolya]

0.4.3 is out (04.11.2026) - the first post and the download links are updated.

• CMP 50HX (Turing) support: the engine is 37% faster on it (22.8 -> 31.3 inf/s, batch 16, core lock 1500 MHz), results byte for byte the node's consensus. The RTX 30 / 40 / 50 engine code is unchanged.
• CMP 50HX numbers (offline bench, batch 16, with the unlock driver): 34.5 inf/s at 219 W (stock), 32.6 inf/s at 151 W (balance), 29.6 inf/s at 134 W (eco).
• HiveOS, on request only: --cmp-unlock-driver in Extra config (or sudo ./ffgv --cmp-driver) installs a driver that unlocks a CMP card; under the stock NVIDIA driver it makes about 11 times less. It is a third-party driver build that replaces the NVIDIA driver of the whole rig and reboots it, so read the manual (section 6) first.
• Auto-undervolt: a search that is cut short keeps its steps, a fault counts only for the GPU with the Xid, a failing start offset is checked once more.

Download: https://github.com/pepsykolya/FreeForGemVela/releases/tag/v0.4.3 (Windows, HiveOS, Linux; SHA256SUMS in the release)

[pepsykolya]

Update: version 0.4.4 is out. The miner now downloads its model itself when it is missing, and AMD Radeon RX 6800 / 6800 XT / 6900 XT / 6950 XT (Navi 21) is supported on Windows. Release and notes are in the first post.

[pepsykolya]

Update: FreeForGemVela 0.4.5 (fixes)

A fix release; the engines (NVIDIA and AMD RX 6800 XT on Windows) are byte for byte those of 0.4.4, so speeds are unchanged.

• Solo mining on your own node reacts to a new block faster. Before, the miner waited for the node to verify a submitted block (the node runs an inference on its CPU, seconds) and did not look for a new tip meanwhile, so all GPUs could work on the old block for 15-17 s (reported by a solo miner). Now submitblock runs in its own thread, a node of your own (local network, Tailscale) is polled every 0.5 s, and the RPC connection is kept (HTTP keep-alive), which also helps on a lossy VPN. Measured on a rig on another site: a new job within 0-1 s of a new block (once 6 s) instead of 4-5 s.
• HiveOS: the dashboard shows "need update!!" next to the version when a newer release is out.
• suprnova is the first pool in all examples; AriaPool is still supported.

Honest note: the sending of a found block was only unit-tested, not on a real node (our test rig has about 0.7% of the network and found no block in 2 hours). Feedback from solo miners is welcome.

Release (Linux / HiveOS / Windows, sha256 on the page): https://github.com/pepsykolya/FreeForGemVela/releases/tag/v0.4.5
HiveOS install URL: https://github.com/pepsykolya/FreeForGemVela/releases/download/v0.4.5/ffgv-hive-0.4.5.tar.gz

[pepsykolya]

Update: FreeForGemVela 0.4.6 - two lanes on RTX 3070, RTX 4070 Ti and CMP 50HX, batch 20 on RTX 3080 Ti

Two lanes (two streams of batches on one GPU: the second fills the idle tails of the first) are now the default on the cards where they were measured to pay off. All figures are the miner's own inf/s (API / log), not pool hashrate:

• RTX 3070 (desktop and laptop): +5..6%
• CMP 50HX: +4% (a 5-card HiveOS rig, 3.3 h on the pool)
• RTX 4070 Ti: +2.4%, a small gain; energy per inference is slightly worse (+1.8%)
• RTX 3080 Ti: automatic batch 20 on one lane instead of 8: +10% on the pool, J/inf -3.8%

• A launch barrier (--lane-sync) keeps the two lanes of an RTX 3070 in phase; without it they can fall into antiphase and run at the speed of one lane (12-19% of the time on the pool).
• The second lane passes the same start checks (test vector, canaries); a hung lane is stopped alone, the GPU goes on; if memory is short the card stays on 1 lane and the log says why.
• --lanes N always wins: --lanes 1 (or --no-lanes) is the 0.4.5 behaviour, --lanes 2 forces two lanes on any card; it works in HiveOS Extra config too. Do not use --lanes 2 with 0.4.5.
• The kernel math is that of 0.4.5, bit-exact with the node's consensus code (test vector, canary 56/56 on every lane).

Honest note: the measurements come from one laptop and one CMP rig; the 4070 Ti gain is small.

Release and files (Windows, Linux, HiveOS; SHA256 on the page): https://github.com/pepsykolya/FreeForGemVela/releases/tag/v0.4.6
Manual: https://github.com/pepsykolya/FreeForGemVela
HiveOS install URL: https://github.com/pepsykolya/FreeForGemVela/releases/download/v0.4.6/ffgv-hive-0.4.6.tar.gz
Video tutorial (EN, 3 parts): 1: the coin | 2: setting up the miner | 3: lanes and undervolt