← All guides

Stock vs custom ASIC firmware: what the dev fee actually buys

Hardware · 9 min read

Firmware is the software that decides how hard each chip in a miner works, at what voltage, under what temperature limits. Stock firmware from the manufacturer runs the machine at one fixed, conservative tune. Custom firmware. Braiins OS+, VNish and LuxOS are the three mature names, promises to tune each chip individually and squeeze out efficiency the factory left behind, in exchange for a percentage of the machine's hashrate, forever. Both sides of that trade are real. This guide prices them from the vendors' own published figures where they exist, flags where the marketing outruns the evidence, and ends with a trial procedure that measures the gain on one machine before any commitment. Nothing here is financial advice.

What firmware actually controls

Three levers matter. Frequency per chip: no two ASIC chips are identical, and a fixed tune must be safe for the weakest chip on the board. Autotuning sweeps each chip at runtime, giving stronger chips more work and weaker chips less. Braiins describes its own version exactly that way. One precision worth keeping, because marketing blurs it: frequency is tuned per chip, but voltage is controlled per power domain, with roughly ten chips sharing a converter. Any claim of per-chip voltage control is an overstatement. Power target: the operator names a wattage ceiling and the firmware maximises hashrate inside it, on Braiins OS, the operator inputs the desired consumption in watts and the firmware calibrates chip frequencies to suit. This is the cleanest way to undervolt for efficiency, or to run a machine gently as a space heater. Protocol and telemetry: among the big three, only Braiins OS+ natively supports Stratum V2, and the custom platforms generally expose far richer per-chip statistics and APIs than stock.

The four realistic options, compared

Stock (manufacturer)Braiins OS+VNishLuxOS
Dev feeNone2 to 2.5% of hashrate (official site)~2 to 2.8% of hashrate~2.8% of hashrate
AutotuningNo, one fixed tuneYes, per-chip frequencyYesYes
Power targetNoYes, watt-levelYesYes
Stratum V2NoYes, nativeNoNo
WarrantyPreservedInstalling third-party firmware generally voids the manufacturer warranty
Source codeClosedCore mining binary closed or only partially open on all three

On the dev fee mechanics, Braiins' own FAQ is the clearest primary source: the fee is 2 to 2.5% depending on the hardware model, collected by directing that percentage of hashing power to the Braiins pool, and it is never exactly the stated percentage in real time because hashrate fluctuates, it stays within a fixed tolerance set out in the licence. Pool pairing changes the total: mining on Braiins Pool with Braiins OS rebates the pool fee to a 0% effective rate instead of the standard 2.5%, with a full refund of the pool fee for a fleet running entirely on Braiins OS. Note what that does and does not do, it waives the pool fee. The dev fee still applies. The other platforms run similar pairings with their own pools, so the honest comparison is always dev fee plus pool fee, together, on the pool the miner will actually use.

The net math: gain minus dev fee, at your power target

Now the part the comparison tables skip. The efficiency claims conflict, and the conflict is instructive. One widely read three-way comparison claims a 10 to 15% improvement in joules per terahash from autotuning, with an S19j Pro moving from a stock 29.5 J/TH to 25 to 28 J/TH. The same publisher's model-specific S19 guide is more conservative: autotuning at stock wattage commonly achieves 27 to 29 J/TH from 29.5, a 2 to 8% improvement, and the larger 15 to 25% efficiency gains appear only in underclocked mode, targeting 2,500 W instead of 3,250 W, which is a different operating point producing less hashrate, not a free lunch at full speed. Treat the first figure as an interested source's best case and the second as the planning figure.

The worksheet is then simple arithmetic, run with the reader's own numbers. Take the conservative gain at the intended power target, say 5% better J/TH at stock watts. Subtract the dev fee at the top of its range, 2.5%. The net edge is roughly 2.5% of energy efficiency, before counting the pool-fee pairing. Run the same worksheet in underclocked mode and the picture flips: a 15 to 25% efficiency gain against the same 2.5% fee is a decisive win if the reduced hashrate still clears the electricity-cost math at the local rate, the trade custom firmware genuinely offers. And run it at the pessimistic end, a 2% gain against a 2.8% fee, and custom firmware is a small net loss bought with a voided warranty. The conclusion the tables never state: the dev fee is not a detail, it is roughly the size of the conservative gain, which is why the trial below measures rather than assumes.

The costs that are not in the tables

Warranty. Installing third-party firmware generally voids the manufacturer warranty, and the manufacturer's own position, as reported from its support material, is that unauthorized firmware causing a malfunction can void coverage. All three platforms can be rolled back to stock, returning to stock Bitmain firmware typically means an SD-card flash with the official recovery image for the model, but rollback restores the software, and a buyer or a warranty desk should not be assumed unable to tell. On a machine already out of warranty the cost is zero, which is why the used market flashes freely; on a machine with ten months of coverage left, the warranty is an asset being spent, and it belongs in the worksheet. Thermal risk. The firmware does not damage hardware by existing, but aggressive overclocking profiles can push chips past their thermal limits where cooling is inadequate; Braiins itself warns that overclocking can damage a device, and that temperature materially changes chip behaviour during tuning. Efficiency profiles run the other way, reducing thermal stress. Auditability. The core binary of the major platforms is closed or only partially open, so the operator cannot fully verify what runs on the machine, the dev fee behaviour has to be observed on the pool dashboard, not read in source. Provenance. The meaningful safety line is between a mature project with documented compatibility, recovery procedures and active support, and an unfamiliar build with none of those. Unknown firmware is also a used-market hazard: it can hide faults or skim hashrate, which is why the buying guide on this site treats unidentifiable firmware as a defect.

The safe trial: one machine, a baseline, a rollback check

Vendor guidance converges on the same shape, return any other custom firmware to stock before installing a new one, and start with a small batch rather than a fleet, but stops short of steps. Here is the full procedure for a single machine:

  1. Record the stock baseline. Run the machine on stock firmware for at least 24 hours at the intended power setting. Log pool-side average hashrate, wall power from a plug meter, and the resulting J/TH. Without this baseline, every later claim is unfalsifiable.
  2. Confirm the exact model and control board against the firmware project's compatibility list, and download the firmware only from the project's official site. Wrong-build flashes are the common avoidable failure.
  3. Save the exit before entering. Download the official stock recovery image for the model and prepare the SD card now, while nothing is on fire.
  4. Flash one machine only, on stable power, and let autotuning complete its sweep before judging anything, tuning is a process with a ramp, not a switch.
  5. Run the identical comparison window: same power target, same pool, at least 24 hours. Log the same three numbers as the baseline, plus the dev-fee behaviour visible on the pool side.
  6. Do the net worksheet with measured figures, measured J/TH gain minus the documented dev fee, and check temperatures stayed inside the stock envelope throughout.
  7. Prove the rollback once: reflash to the stock recovery image, confirm the machine hashes normally, then reflash the custom build if the worksheet justified it. An exit that has never been tested is a hope, not an exit.

What success looks like: a measured efficiency gain that survives subtracting the dev fee, temperatures no worse than the stock baseline, and a demonstrated rollback. A gain visible only in the firmware's own dashboard, with no pool-side confirmation, is not a result.

Choosing: a short decision framework

Under warranty? Stay stock, or price the remaining warranty honestly against a conservative 2 to 8% gain, the trade rarely clears. Out of warranty, running at full power on an expensive rate? The underclocked efficiency mode is the strongest case custom firmware has; model it at the reduced hashrate, not the nameplate. Running one machine as a heater? A watt-level power target is the feature that matters, and any of the three platforms provides it, choose on fee total including the pool pairing. Unverifiable claims, unknown builds, no recovery image? Stock, without apology. The factory tune is boring in the way that roofs are boring: it is the default because the failure modes are known.

FAQ

Is the dev fee charged on top of the pool fee? Yes, they are separate. The dev fee is a percentage of hashrate directed to the firmware maker; the pool fee is the pool's cut of payouts. Some pairings rebate the pool fee to 0% when the matching firmware is used, which is why the combined figure is the only honest comparison.

Can the machine go back to stock? Operationally, yes, the major platforms reflash to stock, typically via an SD-card recovery image for the exact model. Treat the warranty question separately from the software question: reversible software does not guarantee restored coverage.

Which of the three is best? The sources support differences, native Stratum V2 only on Braiins OS+, differing fee ranges, differing pool pairings, but no honest overall winner, because the answer depends on the pool being used and whether the machine runs at full power or underclocked. Measure one machine; the worksheet decides.

Related: Buying a used ASIC: the acceptance test →