Repository navigation
Releases: fearke85/fear-miner
Release list
fear-miner 1.0.2
Upgrade
- HiveOS: set the flight sheet's Installation URL to
fear-miner-1.0.2.tar.gz(link
below). The rest of the flight sheet stays. - Linux and Windows: unpack over the old folder and copy your settings into the new
run.shorstart.bat. Models and escrow files indatastay.
What's new
More Keryx hashrate on RTX 30, RTX 40 and RTX 50 cards, with nothing to set.
Measured on a mining rig at each card's usual overclock, against the previous version:
- RTX 30 series: about +1.0 to +1.7% (RTX 3060 +1.0 to +1.7%, RTX 3070 +1.1%, RTX 3080
+1.6 to +1.7%), at about the same hashrate per watt. - RTX 50 series: about +1.9% on an RTX 5070 Ti locked at 2300 MHz, for 2 to 6 W more.
- RTX 40 series: should gain like RTX 30; not measured yet. Reports are welcome.
- RTX 20 series and CMP 170HX: unchanged.
Downloads
| Platform | File | sha256 |
|---|---|---|
| HiveOS | fear-miner-1.0.2.tar.gz |
6c9c2d41ec0b9c8952505aed0b68035cd8cb462c3c1ab44d4407f7a389d3fd54 |
| Linux | fear-miner-linux-1.0.2.tar.gz |
5d3b5654123ca3e36226035de99def89ad05e63a4e593efc01ef11d5896fed68 |
| Windows | fear-miner-windows-1.0.2.zip |
5e9db5f4f8d3673d9ab3d4e79c7e13079079292988b6dae109ba0b9a7565cd1b |
NVIDIA Turing or newer, driver 580 or newer. Versions before 0.0.22 are invalid since H14.
First install: README (português) ·
Keryx settings: coins/keryx · help on
Discord
fear-miner 1.0.1
Upgrade
- HiveOS: set the flight sheet's Installation URL to
fear-miner-1.0.1.tar.gz(link
below). The rest of the flight sheet stays. - Linux and Windows: unpack over the old folder and copy your settings into the new
run.shorstart.bat. Models and escrow files indatastay.
What's new
- RTX 20 series: back to the hashrate of 0.0.32, and memory overclock raises hashrate again.
1.0.0 lowered hashrate on desktop RTX 20 cards by up to ~13% (reported: RTX 2070 1.42 →
1.24 MH/s). Laptops at their power limit give back the ~4.5% that 1.0.0 gave them. - RTX 40 series: more hashrate, with nothing to set. Measured on an RTX 4090 at stock clocks
with a 410 W power limit: about +2.8% (3.54 → 3.64 MH/s), at about +2% hashrate per watt.
Other RTX 40 cards were not measured yet. Reports of the hashrate you see are welcome. - RTX 30, RTX 50 and CMP 170HX: unchanged from 1.0.0.
Downloads
| Platform | File | sha256 |
|---|---|---|
| HiveOS | fear-miner-1.0.1.tar.gz |
a6994b37193929db9f22b1d1b816c859258d883f6c6da1a6d7245a9212c010bf |
| Linux | fear-miner-linux-1.0.1.tar.gz |
c738fb321560d695f09b569723b2ee066b2714aaa79ce2b5b96f123eea610bf8 |
| Windows | fear-miner-windows-1.0.1.zip |
eb6f4da0ae56f9577bdab2a4563e4fd9a4ab09a92c942544492114be17e4c9ca |
NVIDIA Turing or newer, driver 580 or newer. Versions before 0.0.22 are invalid since H14.
First install: README (português) ·
Keryx settings: coins/keryx · help on
Discord
fear-miner 1.0.0
Upgrade
- HiveOS: set the flight sheet's Installation URL to
fear-miner-1.0.0.tar.gz(link
below). The rest of the flight sheet stays. - Linux and Windows: unpack over the old folder and copy your settings into the new
run.shorstart.bat. Models and escrow files indatastay.
What's new
Higher Keryx hashrate on NVIDIA RTX 20 and RTX 30 cards, with nothing to set. Estimated gain over
0.0.32 by card family, at the same overclock:
- RTX 20 series: about +4.5% (measured on an RTX 2070 Max-Q laptop at its power limit; desktop
cards that do not reach their power limit may gain less). - RTX 30 series: about +1% (measured: RTX 3060 +1.1 to +1.3%, RTX 3070 +1.0%, RTX 3080 +1.0%),
at the same or slightly better hashrate per watt. - RTX 50 series, RTX 40 series and CMP 170HX: unchanged.
Downloads
| Platform | File | sha256 |
|---|---|---|
| HiveOS | fear-miner-1.0.0.tar.gz |
36b476dfc1125bb57d2d2f978daa8d6ed2c301cfd07054fe42afa17c7df36f75 |
| Linux | fear-miner-linux-1.0.0.tar.gz |
d2fa4232b57b28efc1cc0545bb6f25d7c42bd82e9258470e31d0ac54c7c4df2e |
| Windows | fear-miner-windows-1.0.0.zip |
890d5190391bd82ac7635c21f11ece90cf154a5f9808e7db1e283477fcab222d |
NVIDIA Turing or newer, driver 580 or newer. Versions before 0.0.22 are invalid since H14.
First install: README (português) ·
Keryx settings: coins/keryx · help on
Discord
fear-miner 0.0.32 (test)
Upgrade
- HiveOS: set the flight sheet's Installation URL to
fear-miner-0.0.32.tar.gz(link
below). The rest of the flight sheet stays. - Linux and Windows: unpack over the old folder and copy your settings into the new
run.shorstart.bat. Models and escrow files indatastay.
What's new
- NVIDIA CMP 170HX now mines. Cards unlocked to 64 GB mine tier 4. Measured on two unlocked
cards at a 200 W power limit: ~3.2 MH/s per card, 6.4 MH/s for the rig. - No change for other cards.
Downloads
| Platform | File | sha256 |
|---|---|---|
| HiveOS | fear-miner-0.0.32.tar.gz |
030000f41ae66f24a31a82d11acea47667291e782c51b532b494a651f160cf37 |
| Linux | fear-miner-linux-0.0.32.tar.gz |
eae83d92204dd3658df1af104e790957f7742a674690c140028f67763c7f6142 |
| Windows | fear-miner-windows-0.0.32.zip |
21a965644d11ed677350baa71fc81e42093b1a058feea5e98ee0c47d0a515914 |
NVIDIA Turing or newer, driver 580 or newer. Versions before 0.0.22 are invalid since H14.
First install: README (português) ·
Keryx settings: coins/keryx · help on
Discord
fear-miner 0.0.31 (test)
Upgrade
- HiveOS: set the flight sheet's Installation URL to
fear-miner-0.0.31.tar.gz(link
below). The rest of the flight sheet stays. - Linux and Windows: unpack over the old folder and copy your settings into the new
run.shorstart.bat. Models and escrow files indatastay.
What's new
- Cards on narrow PCIe links that mine a tier larger than their memory get much faster, with
nothing to set. They settle about two minutes after start. Measured: RTX 5070 Ti on a 3.0 x1
riser ~120 → 419 kH/s, RTX 3070 on 3.0 x1 117 → 346 kH/s, RTX 5070 Ti on 3.0 x4 ~490 → 876 kH/s. - Fewer strikes after a crash. An AI request the rig still owed when the miner restarted is
answered after the restart while its window is open. - A card that crashes the miner while answering AI requests keeps mining but is kept off AI
requests for a day (then three days, then a week). The next start writes a crash report into
log.txtwith the card, its overclock, temperature, power, PCIe link and the likeliest cause. - Tier 2 requests that arrived while a card was busy mining are no longer dropped.
- Answers use the web results a request carries, as the official miner does.
- Stats API:
--stats ADDRESS:PORTservesGET /stats(default127.0.0.1:4060).run.shand
start.bathave aSTATSline.
Downloads
| Platform | File | sha256 |
|---|---|---|
| HiveOS | fear-miner-0.0.31.tar.gz |
fc64aa1138d63a02f244c7f4a93854b90103a788c025202e9a50a3b2ff97c535 |
| Linux | fear-miner-linux-0.0.31.tar.gz |
6f300bad85812dd07ad61a7a5d5059931abf4b39b4edb4d1277d2f9b91d603fc |
| Windows | fear-miner-windows-0.0.31.zip |
40f899e04a0b882bc503c5c7e81e9baaa75ab08b0990c73d68a238e48e53574a |
NVIDIA Turing or newer, driver 580 or newer. Versions before 0.0.22 are invalid since H14.
First install: README (português) ·
Keryx settings: coins/keryx · help on
Discord
fear-miner 0.0.30 (test)
Test packages. Three archives: HiveOS, plain Linux and Windows.
Tail-copy cards keep more of the walk on the card. A card mining a tier whose
walk does not fit its VRAM (tail-copy mode) now keeps 56 MiB (RTX 3070) to 86 MiB
(RTX 5070 Ti) more of it in VRAM. On narrow risers that is hashrate: the share read
from RAM is what limits these cards. A card that was short of its tier by less than
that now holds the whole walk and mines it at full speed.
Cards that hold their tier mine exactly as before: measured on a mixed rig, beside
two cards in tail-copy mode on tier 3 and one on tier 1, the other four gave the
same hashrate as 0.0.29 (within 0.3%), and the pool accepted every share.
Linux and Windows: this is the first package for them since 0.0.28, so it also
brings 0.0.29's change: every card mines until an AI request arrives, and a request
pauses only the cards it needs.
Still ready for the H14 hard fork (DAA 121,985,000, around 2026-10-09
15:20 UTC), as 0.0.22: older versions produce rejected shares and blocks from that
block on. Solo miners need keryx-node 1.6.4 or newer.
Supported cards: Turing or newer (RTX 20xx, 30xx, 40xx, 50xx) with at least
8 GB, or 6 GB for tier 0 in tail-copy mode beside another card (measured on Ampere
and Blackwell).
Pascal (GTX 10xx, P104/P106) and older
are not supported.
On HiveOS, set the flight sheet's Installation URL to this version's archive (below):
the file name carries the version, and an old name keeps installing the old version.
The rest of the flight sheet stays. On Linux and Windows, copy your settings into the
new run.sh or start.bat.
HiveOS — fear-miner-0.0.30.tar.gz
Miner name stays fear-miner.
/hive/miners/custom/custom-get https://github.com/fearke85/fear-miner/releases/download/v0.0.30/fear-miner-0.0.30.tar.gz
Linux — fear-miner-linux-0.0.30.tar.gz
For a plain box, outside HiveOS. Unpack, set WALLET, WORKER and POOL
(or NODE for solo) at the top of run.sh, and run it:
tar -xzf fear-miner-linux-0.0.30.tar.gz
cd fear-miner
nano run.sh
./run.sh
Windows — fear-miner-windows-0.0.30.zip
Unzip, open start.bat in Notepad, set WALLET, WORKER and POOL (or NODE for
solo), save, and double-click it.
Models, the IPFS repo and escrow go in ./data, so unpacking a new version over
the old one keeps them. Set FEAR_DATA_DIR to put them elsewhere.
All three need an NVIDIA driver new enough for CUDA 13.2 (series 580 or newer). The
archives bring the CUDA libraries; they do not bring the driver.
sha256 c7efa2d409e43a49fa6b5490ebb6e9d4634f60100da3f26ccc435443121a96d3 fear-miner-0.0.30.tar.gz
sha256 3dde9ce383ccff1e7168738b0d47e68aed0ad9505905d2224d4561763c0e8002 fear-miner-linux-0.0.30.tar.gz
sha256 716418943080af613ff681ff47dc502c98f5dcdbb7a4c38fdd18dee0f71255ff fear-miner-windows-0.0.30.zip
fear-miner 0.0.29 (test)
Test package, HiveOS only. Linux and Windows stay on 0.0.28.
Every card mines until an AI request arrives. On rigs that load the AI model
slowly from disk, 0.0.28 took one card out of mining at start (two when the model
is split across a pair) and kept it waiting for an AI request, even when none came.
Now every card mines. A request pauses only the cards it needs, and they go back to
mining once it is answered. When loading the model is slow for the request window,
the log says so at start.
Everything else is as in 0.0.28; see the 0.0.28 release page.
Still ready for the H14 hard fork (DAA 121,985,000, around 2026-10-09
14:00 UTC), as 0.0.22: older versions produce rejected shares and blocks from that
block on. Solo miners need keryx-node 1.6.4 or newer.
Supported cards: Turing or newer (RTX 20xx, 30xx, 40xx, 50xx) with at least
8 GB, or 6 GB for tier 0 in tail-copy mode beside another card (measured on Ampere
and Blackwell).
Pascal (GTX 10xx, P104/P106) and older
are not supported.
On HiveOS, set the flight sheet's Installation URL to this version's archive (below):
the file name carries the version, and an old name keeps installing the old version.
The rest of the flight sheet stays.
HiveOS — fear-miner-0.0.29.tar.gz
Miner name stays fear-miner.
/hive/miners/custom/custom-get https://github.com/fearke85/fear-miner/releases/download/v0.0.29/fear-miner-0.0.29.tar.gz
It needs an NVIDIA driver new enough for CUDA 13.2 (series 580 or newer). The
archive brings the CUDA libraries; it does not bring the driver.
sha256 e18be75a4a5574ab6a684c8431927d3129d3be5b5bebe9f85459fc4ad584d1f3 fear-miner-0.0.29.tar.gz
fear-miner 0.0.28 (test)
Test packages. Three archives: HiveOS, plain Linux and Windows.
Tail-copy mode only where the rig can answer the AI requests. A card in
tail-copy mode mines a tier its VRAM cannot load for AI requests, so those requests
go to another card, or to two cards together. When no card in the rig, alone or
paired with another, can load that tier's model, the requests went unanswered and
the payout took strikes. Such a card now does not mine that tier, and the log and
the TUI say why. Only the cards the miner runs on count: with --device, the
cards left out do not answer AI requests. Examples:
| rig | 0.0.27 | 0.0.28 |
|---|---|---|
one 16 GB card with --tier 3 |
mines tier 3, AI requests unanswered | does not mine; give it a tier it holds |
two 16 GB cards with --tier 3 |
both mine tier 3 | same: the two together answer |
| one 6 GB card alone | mines tier 0, AI requests unanswered | does not mine |
| two or more 6 GB cards | mine tier 0 | same: two together answer |
| a 6 GB card beside a 12 GB card | mines tier 0 | same: the 12 GB card answers |
Cards that hold their tier mine exactly as before.
Everything else is as in 0.0.27: tail-copy mode, --device on HiveOS, Linux
and Windows; see the 0.0.27 release page.
Still ready for the H14 hard fork (DAA 121,985,000, around 2026-10-09
14:00 UTC), as 0.0.22: older versions produce rejected shares and blocks from that
block on. Solo miners need keryx-node 1.6.4 or newer.
Supported cards: Turing or newer (RTX 20xx, 30xx, 40xx, 50xx) with at least
8 GB, or 6 GB for tier 0 in tail-copy mode beside another card (measured on Ampere
and Blackwell).
Pascal (GTX 10xx, P104/P106) and older
are not supported.
On HiveOS, set the flight sheet's Installation URL to this version's archive (below):
the file name carries the version, and an old name keeps installing the old version.
The rest of the flight sheet stays. On Linux and Windows, copy your settings into the
new run.sh or start.bat.
HiveOS — fear-miner-0.0.28.tar.gz
Miner name stays fear-miner.
/hive/miners/custom/custom-get https://github.com/fearke85/fear-miner/releases/download/v0.0.28/fear-miner-0.0.28.tar.gz
Linux — fear-miner-linux-0.0.28.tar.gz
For a plain box, outside HiveOS. Unpack, set WALLET, WORKER and POOL
(or NODE for solo) at the top of run.sh, and run it:
tar -xzf fear-miner-linux-0.0.28.tar.gz
cd fear-miner
nano run.sh
./run.sh
Windows — fear-miner-windows-0.0.28.zip
Unzip, open start.bat in Notepad, set WALLET, WORKER and POOL (or NODE for
solo), save, and double-click it.
Models, the IPFS repo and escrow go in ./data, so unpacking a new version over
the old one keeps them. Set FEAR_DATA_DIR to put them elsewhere.
All three need an NVIDIA driver new enough for CUDA 13.2 (series 580 or newer). The
archives bring the CUDA libraries; they do not bring the driver.
sha256 d0b267542edfc6dd6169a3b0b0cb746d32ff27b5ed8a1793014f26c9a1505e95 fear-miner-0.0.28.tar.gz
sha256 9cc6be57d6c4260bee6a207994bc4e111fa7a5fd4a5e03c94ad864e62d707102 fear-miner-linux-0.0.28.tar.gz
sha256 713295dedab03db4d8fcfaf7e358823ae170dc717ce73136717af73e0e05af2e fear-miner-windows-0.0.28.zip
fear-miner 0.0.27 (test)
Test packages. Three archives: HiveOS, plain Linux and Windows.
Choose which cards mine with --device. List the cards by the numbers the
miner shows (gpu 0, gpu 1 ... in the log and the TUI); --tier then gives
one tier per listed card, in the same order. On HiveOS put it in the flight sheet's
Extra config arguments, before the --: --device 1,6 -- --tier 3,3. Until now
a --device there stopped the miner from starting. On Linux set
DEVICES="1,6" in run.sh, on Windows set "DEVICES=1,6" in start.bat;
copy your settings into the new script.
Everything from 0.0.23 to 0.0.26:
Windows: tail-copy mode now uses all of the card's VRAM. On Windows the driver
reports less free VRAM than the card really holds, and tail-copy cards sized
themselves by that number: up to about 0.8 GB of VRAM sat idle while that much more
of the model was read from RAM. They now measure what fits. Measured on an RTX
5070 Ti under Windows with no monitor, tier 3: 95% to 97.8% of the model in VRAM,
and the RAM it takes from 0.8 GB to 0.36 GB.
The miner version shows in the TUI header and on every log line, so a log
excerpt says which version wrote it.
Tail-copy mode is faster on every PCIe link and uses less RAM. A card in
tail-copy mode now keeps more of the model in VRAM on any link, not only on risers,
so less of it crosses PCIe on each hash. Measured on two RTX 5070 Ti mining tier 3
(Qwen3.6-27B) on 16 GB: on PCIe 5.0 x4, 1.63 to 2.68 MH/s; on 5.0 x8, 2.60 to
2.62 MH/s. And every tail-copy card mining the same model now shares one copy of
the part kept in RAM: that rig went from 612 + 490 MB of RAM to 345 MB.
One card the AI engine cannot run on no longer stops the miner. At start the
miner checks that each card can run AI requests. A card that failed that check
used to stop the whole rig; now that card keeps mining, the other cards answer the
AI requests, and the log names the card and the reason. The miner only refuses to
start when no card passes, or when the card given in --inference-gpu fails.
6 GB cards now mine tier 0. A card whose VRAM cannot hold the tier-0 model
(RTX 3050 6 GB, laptop RTX 3060 6 GB, an 8 GB card whose monitor left it under
tier 0) used to mine nothing. It now mines tier 0 in
tail-copy mode: the part of the model that does not fit in VRAM stays in system
RAM and is read over PCIe. Cards that hold their tier keep mining exactly as
before; this mode never turns on for them by itself.
RAM. A 6 GB card needs about 0.5 GB of RAM for the part left out of VRAM, and
the miner keeps 1 GB free for the system on top of that: have at least 2 GB of
RAM free when the miner starts (4 GB in the rig). All tail-copy cards on the same
model share one copy, sized for the card that leaves the most out. When the RAM is not there, that card
does not start, the log says how much it needs and how much is free, and the
other cards mine as usual.
The faster the PCIe link, the more hashrate. In this mode the card's link
(riser or slot) sets the speed: a newer PCIe generation and more lanes give more
hashrate. Measured on tier 0 with a 6 GB card's share of the model in RAM:
| link | hashrate |
|---|---|
| PCIe 3.0 x1 riser | ~45 kH/s |
| PCIe 3.0 x4 | ~180 kH/s |
A PCIe 2.0 x1 riser gives about half of a 3.0 x1. On PCIe 5.0 x4 and x8, cards a few
hundred MB short of their tier measured at their usual hashrate. A card only a
little short of its tier loses little: 9 MB short of tier 0, a 3060 measured 0.47 MH/s on a 3.0 x1 riser and
1.01 MH/s on 3.0 x4 (1.04 MH/s with the whole model in VRAM).
Forcing a higher tier. --tier sets the tier of every card (--tier 1) or
of each card in --device order (--tier 2,0,2). A card forced above what its
VRAM holds now mines that tier in tail-copy mode, with the same RAM and PCIe rules,
instead of failing. It is slower than a tier the card holds, and much slower on a
riser. On HiveOS put it in the flight sheet's Extra config arguments
(--tier 2,0,2); on Linux in EXTRA=(--tier 2,0,2) in run.sh; on Windows in
set "EXTRA=--tier 2,0,2" in start.bat.
Still ready for the H14 hard fork (DAA 121,985,000, around 2026-10-09
14:00 UTC), as 0.0.22: older versions produce rejected shares and blocks from that
block on. Solo miners need keryx-node 1.6.4 or newer.
Supported cards: Turing or newer (RTX 20xx, 30xx, 40xx, 50xx) with at least
8 GB, or 6 GB for tier 0 in tail-copy mode (measured on Ampere and Blackwell).
Pascal (GTX 10xx, P104/P106) and older
are not supported.
On HiveOS, set the flight sheet's Installation URL to this version's archive (below):
the file name carries the version, and an old name keeps installing the old version.
The rest of the flight sheet stays. On Linux and Windows, copy your settings into the
new run.sh or start.bat.
HiveOS — fear-miner-0.0.27.tar.gz
Miner name stays fear-miner.
/hive/miners/custom/custom-get https://github.com/fearke85/fear-miner/releases/download/v0.0.27/fear-miner-0.0.27.tar.gz
Linux — fear-miner-linux-0.0.27.tar.gz
For a plain box, outside HiveOS. Unpack, set WALLET, WORKER and POOL
(or NODE for solo) at the top of run.sh, and run it:
tar -xzf fear-miner-linux-0.0.27.tar.gz
cd fear-miner
nano run.sh
./run.sh
Windows — fear-miner-windows-0.0.27.zip
Unzip, open start.bat in Notepad, set WALLET, WORKER and POOL (or NODE for
solo), save, and double-click it.
Models, the IPFS repo and escrow go in ./data, so unpacking a new version over
the old one keeps them. Set FEAR_DATA_DIR to put them elsewhere.
All three need an NVIDIA driver new enough for CUDA 13.2 (series 580 or newer). The
archives bring the CUDA libraries; they do not bring the driver.
sha256 2b10f29762c4f40ab1dc6c7cf2ce70d86897dd378f8884df46fa81f5414a0785 fear-miner-0.0.27.tar.gz
sha256 ee0a7ccbdb28a3f142a12bc40dd81d1a7f864e6180badbf9d4e6a73b7115425f fear-miner-linux-0.0.27.tar.gz
sha256 619c57cd0ed81f818780a1ec13d4bcd6c2b0f6fd0be91e33660cb52ad705f855 fear-miner-windows-0.0.27.zip
fear-miner 0.0.26 (test)
Test packages. Three archives: HiveOS, plain Linux and Windows.
Windows: tail-copy mode now uses all of the card's VRAM. On Windows the driver
reports less free VRAM than the card really holds, and tail-copy cards sized
themselves by that number: up to about 0.8 GB of VRAM sat idle while that much more
of the model was read from RAM. They now measure what fits. Measured on an RTX
5070 Ti under Windows with no monitor, tier 3: 95% to 97.8% of the model in VRAM,
and the RAM it takes from 0.8 GB to 0.36 GB.
The miner version shows in the TUI header and on every log line, so a log
excerpt says which version wrote it.
Everything from 0.0.25, 0.0.24 and 0.0.23:
Tail-copy mode is faster on every PCIe link and uses less RAM. A card in
tail-copy mode now keeps more of the model in VRAM on any link, not only on risers,
so less of it crosses PCIe on each hash. Measured on two RTX 5070 Ti mining tier 3
(Qwen3.6-27B) on 16 GB: on PCIe 5.0 x4, 1.63 to 2.68 MH/s; on 5.0 x8, 2.60 to
2.62 MH/s. And every tail-copy card mining the same model now shares one copy of
the part kept in RAM: that rig went from 612 + 490 MB of RAM to 345 MB.
One card the AI engine cannot run on no longer stops the miner. At start the
miner checks that each card can run AI requests. A card that failed that check
used to stop the whole rig; now that card keeps mining, the other cards answer the
AI requests, and the log names the card and the reason. The miner only refuses to
start when no card passes, or when the card given in --inference-gpu fails.
6 GB cards now mine tier 0. A card whose VRAM cannot hold the tier-0 model
(RTX 3050 6 GB, laptop RTX 3060 6 GB, an 8 GB card whose monitor left it under
tier 0) used to mine nothing. It now mines tier 0 in
tail-copy mode: the part of the model that does not fit in VRAM stays in system
RAM and is read over PCIe. Cards that hold their tier keep mining exactly as
before; this mode never turns on for them by itself.
RAM. A 6 GB card needs about 0.5 GB of RAM for the part left out of VRAM, and
the miner keeps 1 GB free for the system on top of that: have at least 2 GB of
RAM free when the miner starts (4 GB in the rig). All tail-copy cards on the same
model share one copy, sized for the card that leaves the most out. When the RAM is not there, that card
does not start, the log says how much it needs and how much is free, and the
other cards mine as usual.
The faster the PCIe link, the more hashrate. In this mode the card's link
(riser or slot) sets the speed: a newer PCIe generation and more lanes give more
hashrate. Measured on tier 0 with a 6 GB card's share of the model in RAM:
| link | hashrate |
|---|---|
| PCIe 3.0 x1 riser | ~45 kH/s |
| PCIe 3.0 x4 | ~180 kH/s |
A PCIe 2.0 x1 riser gives about half of a 3.0 x1. On PCIe 5.0 x4 and x8, cards a few
hundred MB short of their tier measured at their usual hashrate. A card only a
little short of its tier loses little: 9 MB short of tier 0, a 3060 measured 0.47 MH/s on a 3.0 x1 riser and
1.01 MH/s on 3.0 x4 (1.04 MH/s with the whole model in VRAM).
Forcing a higher tier. --tier sets the tier of every card (--tier 1) or
of each card in --device order (--tier 2,0,2). A card forced above what its
VRAM holds now mines that tier in tail-copy mode, with the same RAM and PCIe rules,
instead of failing. It is slower than a tier the card holds, and much slower on a
riser. On HiveOS put it in the flight sheet's Extra config arguments
(--tier 2,0,2); on Linux in EXTRA=(--tier 2,0,2) in run.sh; on Windows in
set "EXTRA=--tier 2,0,2" in start.bat.
Still ready for the H14 hard fork (DAA 121,985,000, around 2026-10-09
14:00 UTC), as 0.0.22: older versions produce rejected shares and blocks from that
block on. Solo miners need keryx-node 1.6.4 or newer.
Supported cards: Turing or newer (RTX 20xx, 30xx, 40xx, 50xx) with at least
8 GB, or 6 GB for tier 0 in tail-copy mode (measured on Ampere and Blackwell).
Pascal (GTX 10xx, P104/P106) and older
are not supported.
On HiveOS the flight sheet does not change. On Linux and Windows, copy your
settings into the new run.sh or start.bat.
HiveOS — fear-miner-0.0.26.tar.gz
Miner name stays fear-miner.
/hive/miners/custom/custom-get https://github.com/fearke85/fear-miner/releases/download/v0.0.26/fear-miner-0.0.26.tar.gz
Linux — fear-miner-linux-0.0.26.tar.gz
For a plain box, outside HiveOS. Unpack, set WALLET, WORKER and POOL
(or NODE for solo) at the top of run.sh, and run it:
tar -xzf fear-miner-linux-0.0.26.tar.gz
cd fear-miner
nano run.sh
./run.sh
Windows — fear-miner-windows-0.0.26.zip
Unzip, open start.bat in Notepad, set WALLET, WORKER and POOL (or NODE for
solo), save, and double-click it.
Models, the IPFS repo and escrow go in ./data, so unpacking a new version over
the old one keeps them. Set FEAR_DATA_DIR to put them elsewhere.
All three need an NVIDIA driver new enough for CUDA 13.2 (series 580 or newer). The
archives bring the CUDA libraries; they do not bring the driver.
sha256 bfba25707931a68f7957e83698c2faa49538dd1961482f77677bbad77be280ab fear-miner-0.0.26.tar.gz
sha256 291dbd634dc7880f36adfcac1faa238dc309ca30140efea1f3fdb071c004c320 fear-miner-linux-0.0.26.tar.gz
sha256 253410705f3932e1dedb250a08cc443bb6a23941f3df0dd5ede857250ca17f13 fear-miner-windows-0.0.26.zip