Guide
Overclocking, fan control and safety limits
Power limits, clock offsets and locks, fan curves, per-card settings, and what to reach for per algorithm.
There are three ways to overclock in bzminer — a start script or config.txt, the
console, and the web dashboard — and they all go through one implementation, so they cannot
disagree with each other. Every refusal tells you why (“needs root/administrator”,
“not supported by this card”) instead of failing quietly. This guide is the whole of it.
What each device type can do
| Device | Power | Core clock | Memory clock | Fan |
|---|---|---|---|---|
| NVIDIA | yes | offset and lock | offset and lock | duty, temperature target, or automatic |
| AMD | — | — | — | duty, temperature target, or automatic |
| Intel | — | — | — | — |
| CPU | — | — | — | — |
NVIDIA is the full set. On AMD only the fan is settable; ask for a clock or a power limit on a
Radeon and bzminer says so by name rather than reporting a card that is not there. Intel cards
report their fan but expose no way to drive it. There is no CPU overclocking; use
--cpu_threads to decide how much of the processor mines instead.
On Linux every knob is a privileged driver write. Run as root or nothing applies. On AMD, fan control additionally needs the amdgpu overdrive gate, which bzminer turns on at startup when it is root. On Windows the power limit and fan need no special privileges. bzminer restores whatever it changed when it exits normally.
From the command line, or config.txt
--oc-power-limit 160 board power cap, watts
--oc-core-clock-offset 150 core offset, MHz
--oc-memory-clock-offset 1000 memory offset, MHz
--oc-lock-core-clock 1600 pin the core clock; 0 unlocks
--oc-lock-memory-clock 810 pin the memory clock; 0 unlocks
--oc-fan-speed 60 fan duty, or a temperature target (below)
--oc-reset back to driver defaults
Each takes one value for every card, or one per card: --oc-power-limit 160,180
sets 160 W on the first and 180 W on the second, and a short list leaves the remaining
cards untouched. A worked example — XELIS with the cards capped, a mild core offset, memory
pushed harder, and the fans on a temperature target:
./bzminer -a xelis -p stratum+ssl://us.vipor.net:5177 -w xel:YOUR_WALLET \
--oc-power-limit 160 --oc-core-clock-offset 150 --oc-memory-clock-offset 1000 \
--oc-fan-speed "t:60[25-75] tm:80[50-100]"
The underscore spellings from bzminer 1.x (--oc_power_limit, --oc_fan_speed)
are accepted too, so an old start script needs no editing.
One card at a time
The devices block of config.txt holds per-card settings. bzminer fills in
which card each entry is on its first run — PCI address, name, architecture, board id — so
you only find the one you want and fill in the fields you care about:
{
"oc": { "power_limit": "160", "fan_speed": "t:60[25-75]" },
"devices": [
{ "pci": "0000:07:00", "name": "NVIDIA GeForce RTX 5060 Ti", "core_clock_offset": "150", "memory_clock_offset": "1000" },
{ "pci": "0000:09:00", "name": "NVIDIA GeForce RTX 3060", "core_clock_offset": "-100", "power_limit": "130" },
{ "pci": "0000:03:00", "name": "gfx1100", "fan_speed": "70" }
]
}
Anything left blank falls back to the global oc block, so the above caps the first two
cards at 160 W except the second, which asks for 130 and gets it. Entries are matched on
pci, which does not move when a card is added or removed.
Offsets and locks are different things
An offset shifts the card’s whole voltage/frequency curve and it still boosts. A
lock pins the clock. A card can carry both, which is why each has its own flag and its own
hotkey. A locked memory clock is only accepted at one of the board’s own supported
states — ask for anything else and the driver quietly runs the nearest one it supports, so
bzminer reads the clock back and tells you what the card is actually doing. The device page
(d in the console) lists a card’s lockable memory clocks, with the current one in
brackets.
Fan control
--oc-fan-speed 60 just run at 60%
--oc-fan-speed "t:60[25-75]" hold the CORE at 60C, using 25-75% fan
--oc-fan-speed "th:70[30-90]" the same, aimed at the HOTSPOT
--oc-fan-speed "tm:80[50-100]" the same, aimed at MEMORY temperature
--oc-fan-speed "t:60[25-75] tm:80[50-100]" both - whichever wants more fan winsThat last form is the one worth knowing. A card can sit at a perfectly comfortable core temperature while its memory cooks, and the second clause is what catches it. The fan moves about a point per second toward the target rather than jumping, and stays inside the range you gave. It works on NVIDIA and AMD, because the curve is computed by bzminer and handed to the driver as a duty.
A card keeps a fan duty or a locked clock until something clears it. bzminer puts back everything it set when it stops - Ctrl-C, closing its window, a mining OS stopping it - but a run that is force-killed (kill -9, taskkill /F, a crash) leaves them behind; bzminer --oc-reset on its own puts every card back to driver defaults.
Live, from the console
Press c to open the command line, then:
oc what is there, and what each card is set to
oc 150 500 +150 core, +500 memory on every card
oc 0 core=+150 mem=+500 card 0 only
oc all fan=70 every card
oc 1 power=300 clock=2600 card 1: 300 W limit, 2600 MHz locked core
oc 2 mem=1200 fan=85 card 2: +1200 memory, fans at 85%
oc 3 memclock=810 card 3: memory locked to 810 MHz
oc 0 fan=auto hand the fan back to the driver
oc reset [dev|all] undo it
The settings are core, mem, power, clock,
memclock and fan. core and mem are offsets in MHz;
clock and memclock lock to an absolute MHz, with 0 unlocking.
A device is named by its number, by its PCI address as the tables print it
(oc 29:0 fan=70), or by all. fan=auto is not
fan=0: zero means stop the fan.
With the hotkeys
For nudging by feel while watching what a number does. They move one card, the one n is aimed at, marked > in the tables:
| Key | Step |
|---|---|
g / G | core clock offset, ±15 MHz |
l / L | core clock lock, ±15 MHz |
m / M | memory clock offset, ±50 MHz |
k / K | memory clock lock, to the card’s next / previous supported state |
w / W | board power cap, ±5% of the card’s range |
n | aim the keys at the next card |
Each press starts from what the card is actually running, read back per device, so a rig
started with an offset in config.txt does not jump on the first press, and the result
line says which knob moved and where it landed.
From the dashboard
Each GPU card in the web dashboard has a Tune button when the card is controllable. The panel offers only the knobs the card supports, with the ranges the driver reported, and posts to the same service the console uses. The reply is green, amber or red for all, some, or none of the requested knobs landing, with the driver’s reason shown verbatim.
What to reach for, per algorithm
Starting points, not measured claims — every card is different, and the honest test is
--benchmark at a fixed difficulty,
comparing accepted shares.
| Algorithm | Bound by | What usually pays |
|---|---|---|
| ergo | memory bandwidth | memory offset first, then trim the power limit — the core has slack |
| ethash, etchash | memory bandwidth, almost entirely | memory offset first, then take the power limit down hard — the core spends most of its time waiting on DAG reads |
| kawpow | memory, with a real core component | memory offset, and a core lock rather than a big offset: KawPow’s power draw is spiky and boost wanders |
| xelis | memory | memory offset, power limit down; the core buys little |
| pearl | compute | core clock, and headroom: the prover is heavy, so do not starve it on power |
| cn/gpu | compute | a small core offset and a power cap found by measurement; the classic variants want the memory clock instead |
| warthog | both halves at once | leave the GPU near stock and give the CPU room — a card pushed until it throttles drags the whole pair down |
| randomx, verus | CPU cache and memory | nothing here applies; see --cpu_threads |
Two rules that hold everywhere: cap power before you raise clocks — an efficiency gain
is usually a watt saved, not a megahertz added — and change one thing at a time,
because poolhr takes hours to mean anything and two changes at once cannot be told apart.
Safety limits
Off by default. Turn on what you want and bzminer parks a card that misbehaves and brings it back when it recovers — only the offending card stops; the rest of the rig keeps mining.
--set safety.max_temp_c=85 # pause a GPU over 85 C
--set safety.max_power_w=600 # ...or over 600 W board power
--set safety.sustain_s=15 # only after 15s over the limit, not on a spike
--set safety.resume_margin=5 # resume once it is 5 back under
--set safety.resume_s=10 # ...and has stayed there 10s
Set the temperature limit first. It works on every card, and hardware in trouble — a failing
12VHPWR connector included — gets hot. There are 12VHPWR current limits too
(safety.max_current_a, and per-pin safety.max_pin_current_a and
safety.max_pin_imbalance_a), but they need sensors only some boards carry;
config.txt says exactly which.
Problems
- Nothing applies on Linux. The miner is not running as root. Start it with
sudo. - A memory lock lands on a different number. Locks are only honoured at the board’s own states; bzminer tells you which one you got. The device page lists them.
- The card still has a fan duty after a crash. Run
bzminer --oc-reseton its own to put it back. - An AMD card refuses a clock or power limit. That is a limit of what bzminer can control on Radeon rather than a fault: only the fan is settable there.