The BM1373 Era: Home Mining Goes 3nm — Complete Reference

In November 2025 the fastest chip a home miner could buy was the BM1370: 5nm, roughly 1.2 TH/s per chip, about 15 J/TH. Eight months later a desktop box draws 400 watts and produces 30 TH/s. The chip that did it is the BM1373, and this article is the complete reference on the era it started: every machine built on it, what the numbers actually mean, and the one calculation that decides whether any of it makes sense where you live.

What is the BM1373 era?

The BM1373 era is the generation of home Bitcoin miners built around Bitmain’s 3nm SHA-256 ASIC, which reached desktop users in mid-2026. The chip’s published specification is approximately 2.5 TH/s per unit at about 25 W, or roughly 10 J/TH.

Two supply-chain facts shape every product in this era, and neither appears on a spec sheet.

The chips are harvested, not bought. Early BM1373 supply comes from salvaged Antminer S23 and S23 Hydro hashboards rather than a Bitmain retail channel. Reported developer sourcing cost sits near 60 US dollars per chip. A four-chip miner therefore carries about 240 dollars of raw silicon before enclosure, power supply, controller, certification or margin, which is why the entry price of this generation jumped so sharply over the BM1370 era.

Early yield is inconsistent. Two chips from the same batch can differ materially in the maximum frequency they hold without producing invalid hashes. This is the ordinary silicon lottery, but it is amplified on a brand-new process node fed by salvage parts. It is the single most important reason to tune by measured error rate rather than by copying settings from a forum, and we cover the method in the configuration section below.

One vendor, TinyChipHub, sells chips in sealed reels from a direct channel rather than salvage. That distinction matters more than any marketing claim on the page: a fresh chip has no prior thermal history.

What are the five BM1373 machines?

Read that as two markets, not one. Single-chip units are efficiency instruments: they exist to put a modern chip on a desk at the lowest possible running cost. Multi-chip units are hashrate instruments: they maximise lottery tickets per dollar and pay for it in watts.

NerdAxe Gaia — the open-source reference point

Direct descendant of the Bitaxe and NerdAxe lineage. One BM1373, 2.4 TH/s, 25 W, about 10.4 J/TH, 399.99 US dollars with power supply, currently pre-order with a 6 to 8 week lead time. Vendor imagery is still an approximated mock-up, which tells you the unit was announced ahead of production.

Choose it for firmware transparency. The AxeOS lineage is community-maintained, you can flash it yourself, and you can verify exactly where your hashrate goes. For anyone whose reason for solo mining is sovereignty rather than expected value, that property is the entire product. Do not choose it on economics: 167 dollars per terahash is the worst ratio in the group, roughly 6 times the BSB B30.

Zyber Blanc — the efficiency champion

The most interesting unit on measured numbers rather than claimed ones. One BM1373 in a 135 by 60 by 50 mm screenless enclosure with programmable RGB status lighting, discounted to 299 dollars from 399, powered by a 5 V / 6 A external brick.

The vendor claims up to 3.3 TH/s at 20 W and higher, in the 8 to 10 J/TH band. More usefully, the vendor published a bench measurement: 8.70 J/TH peak, 8.56 J/TH average, at 15.2 W system draw. Do the arithmetic and that test was running near 1.75 TH/s, not 3.3. It is an excellent efficiency figure at a conservative operating point, not a 3.3 TH/s efficiency figure. Read every published J/TH number this way, including the ones in the table above.

Firmware is mid-migration: units shipped on an AxeOS transition layer while the vendor’s in-house Zyber OS finished development. Zyber OS promises deeper BM1373 voltage and frequency curve optimisation, but it is not community firmware.

Nexus S1 — the balanced flagship

Four BM1373 chips, 10.0 to 10.3 TH/s, 100 to 103 W, AXP90 full-copper heatsink with heat pipe, dual front and rear fans, automotive-grade fuse protection, roughly 50 dB, XT30 power. Connectivity varies by reseller: some list 2.4 GHz WiFi via an ESP32-S3-WROOM-1, others Ethernet, some both. Confirm before ordering.

Price is where it gets confusing. The same machine has been listed at 599 dollars direct, 699 at HeliumDeploy and 780 at XC Miner. That is a 30 percent spread on an identical product, so shop it aggressively.

It is the only unit that hits the chip’s rated 10 J/TH at meaningful scale while keeping fully open firmware and a 100 W wall-plug envelope. That is the best compromise position on the board. One caution: early units have been reported drawing 120 to 140 W rather than 100 W.

Nexus S1-1 — the spec that does not add up

Marketed as the single-chip little brother: 6 TH/s at 95 W, rear-fan cooling, integrated Ethernet.

Those numbers imply about 15.8 J/TH. That is worse than the previous-generation BM1370 and 58 percent worse than every other single-chip BM1373 unit, which land at 2.4 to 3.3 TH/s and 20 to 25 W. Either the listing contains an error, or the chip is being driven at an overclock so aggressive it discards the entire reason to buy a 3nm part. One customer review notes the unit runs at roughly 40 percent the speed of visually identical hardware using a faster chip package. Verify with the seller before spending money.

BSB B30 — the cheapest terahash, with two catches

Six chips, approximately 30 TH/s, 360 to 450 W, 799 dollars. At 26.63 dollars per terahash it is the cheapest hashrate in home mining by a wide margin, roughly 2.6 times cheaper per TH than the Nexus S1.

The engineering is a genuine step up from reference-design rebadges. A 94 percent efficient power supply is built in behind a standard IEC C14 socket, so any computer power cord works. There is a replaceable built-in fuse, two fans, a colour display, a programmable front LED strip, an onboard buzzer that can sound on block-found and warning events, and both WiFi and Ethernet. It accepts 120 V and 220 V. Batches 1 and 2 sold out; batch 3 ships 30 September 2026.

Catch one is firmware. BSB OS is closed source. The vendor’s stated reasoning is one accountable official build, with no mystery factory images and no reseller-modified binary quietly redirecting hashrate while the dashboard looks fine. That is a real problem in this market and the argument is legitimate. But it means you cannot audit where your shares go, you cannot flash a community build without voiding warranty, and the fleet-management feature that lets BSB OS also control your Bitaxes and NerdAxes concentrates a great deal of trust in one vendor.

Catch two is efficiency, and it costs more than most buyers realise. See the next two sections.

The efficiency reality check

Vendors across this generation advertise up to 50 percent less power than the BM1370. The honest generational figure is about 33 percent: roughly 15 J/TH becomes roughly 10 J/TH. Only well-cooled single-chip and four-chip designs actually reach the chip’s rating.

Claimed vs derived efficiency (derived = rated power / rated hashrate)

  Zyber Blanc   claimed  8-10 J/TH   measured  8.70 J/TH  @ 1.75 TH/s
  Nexus S1      claimed  ~10 J/TH    derived  10.00 J/TH  (103W / 10.3TH)
  NerdAxe Gaia  claimed  ~10 J/TH    derived  10.42 J/TH  (25W / 2.4TH)
  BSB B30       claimed  12-15 J/TH  derived  13.33 J/TH  (400W / 30TH)
  Nexus S1-1    claimed  none        derived  15.83 J/TH  (95W / 6TH)

  BM1370 generation reference:                ~15.00 J/TH

The B30 sits closer to the previous generation than to its own chip’s rating. That is not a defect, it is a design decision: six chips pushed hard in one box optimises dollars per terahash, not joules per terahash. But it has a consequence that no product page states.

The break-even electricity law

Here is the calculation that should drive every purchase in this era, and the one we have not found published anywhere for home miners.

The electricity price at which a miner stops being expected-value positive depends only on its efficiency in J/TH. Not on its hashrate. Not on what you paid for it. Not on how many chips it has.

The reason is arithmetic. Revenue scales linearly with hashrate. Power consumption also scales linearly with hashrate. Divide one by the other and hashrate cancels out, leaving efficiency and the network’s hashprice.

Break-even electricity price ($/kWh)
  = hashprice_per_TH_per_day / (efficiency_J_per_TH x 0.024)

where hashprice_per_TH_per_day = network hashprice / 1000

At the 11 July 2026 hashprice of $31.10 per PH/s per day:

Three conclusions follow, and they reframe the whole comparison.

The whole generation lives inside a narrow band, roughly 0.08 to 0.15 dollars per kWh. Below 0.08 every machine here is expected-value positive and the only question is acquisition cost. Above 0.15 none of them are, and the machine becomes a lottery ticket plus a hobby, which is a legitimate purchase but should be priced as one. The interesting decisions all happen in between, and that is exactly where most residential rates sit.

What that costs in practice, at 24/7 operation:

Note how differently the two halves of the lineup behave across that range. A single-chip unit goes from pocket change to slightly more pocket change: the difference between the cheapest and most expensive tier is under 5 dollars a month, which is not a decision. The B30 swings by 72 dollars a month across the same range, which is a decision. The higher your electricity price, the more the choice of machine matters, and the more it tilts toward the small ones.

There is also a heat consideration no spec sheet lists. 400 W of continuous dissipation is noticeable in a small room, welcome in a cold month and unwelcome in a hot one. Our hashrate-as-heating analysis covers when that offsets the power bill and when it adds an air-conditioning bill on top of it.

The B30’s cost advantage inverts once the power bill starts. It is 2.6 times cheaper to acquire per terahash and it has the second-worst break-even price in the group. Those two facts are the same fact seen from opposite ends of the ownership period. Below roughly 0.09 dollars per kWh the B30 is the obvious choice. Above roughly 0.15 it is the worst choice in the lineup.

Efficiency is the only spec that ages well. Hashprice falls over time as difficulty rises and after each halving. When it falls, every break-even figure in that table falls with it, proportionally. The most efficient machine is the last one still worth running, and it will still be the last one five years from now.

Multi-chain solo odds: the numbers nobody publishes

Every comparison of these machines quotes Bitcoin odds and stops. But the BM1373 is a SHA-256 chip, and SHA-256 secures several chains whose difficulty is orders of magnitude below Bitcoin’s. The hardware does not care which one it points at. Only the difficulty changes.

Reference values as of 18 July 2026: Bitcoin difficulty 127.17 T at roughly 891 EH/s; Bitcoin Cash difficulty 428.92 G at roughly 2.84 EH/s; eCash at roughly 91.81 PH/s, from which difficulty derives to approximately 12.8 G. Mean time to a block is difficulty times 2 to the 32nd, divided by your hashrate.

That last row is the headline of this entire article. A 799 dollar box on a desk finds an eCash block roughly every three weeks and a Bitcoin Cash block roughly every two years. Those are not lottery odds. Those are odds you can plan around.

Now the honest other half. A Bitcoin Cash block at 3.126 BCH is worth roughly 730 US dollars at July 2026 prices. An eCash block at 1,812,500 XEC is worth roughly 10 dollars. Frequent blocks on a low-value chain are not the same thing as income, and any comparison that shows you the odds without the block value is selling you something.

What the eCash column actually buys you is feedback. Twenty-one days between blocks means you can verify your configuration works end to end, watch a real coinbase transaction land in your own wallet, and confirm your payout address is correct, all within a month. On Bitcoin you would wait five centuries to learn the same thing. For anyone testing a setup before committing it to a serious chain, that is worth far more than ten dollars.

BC2 and BCH2 sit between these extremes with their own difficulty and reward profiles. Difficulty on smaller chains moves faster in relative terms than Bitcoin’s, so treat every number here as a snapshot and pull live figures from the solo mining calculator or the network radar before deciding. Strategy for choosing a chain is covered in solo versus pool mining.

Fleet arithmetic: what 30 TH/s costs five different ways

A useful way to see the trade-offs is to hold hashrate constant and vary the machine.

The B30 is 6 times cheaper than the equivalent Gaia fleet and uses 88 W more. That is the entire argument for and against it in one line. A distributed fleet also gives you redundancy, independent failure domains and per-unit tuning; a single box gives you one power cord, one IP address and one thing to fail.

Configuration: getting any BM1373 miner running properly

Everything below applies to units running AxeOS or an AxeOS-derived interface, which covers the NerdAxe Gaia, both Nexus S1 variants and Zyber Blanc units still on the transition layer. The BSB B30 exposes similar concepts through its own dashboard but does not accept community firmware.

The tuning method that actually works

Do not copy frequency and voltage numbers from a forum post. Early BM1373 silicon varies enough between units that another person’s stable point may be your unstable one. Tune by measured error rate:

1. Record the stock frequency and core voltage before changing anything. Write them down somewhere outside the device. This is your recovery path.
2. Run 30 minutes untouched. Note baseline hardware error percentage and chip temperature.
3. Raise frequency by one 25 MHz step. Leave voltage alone.
4. Run 30 minutes. If error rate stays below 2 percent and temperature is stable, repeat step 3.
5. When error rate crosses 2 percent, back frequency down 25 MHz. That is your chip’s stable ceiling at this voltage.
6. Only then consider raising core voltage in 25 mV steps to unlock further frequency, accepting the power and heat cost.
7. Run 24 hours at the final setting before calling it stable. Many failures appear only after full thermal soak.

Frequency-to-hashrate is close to linear: 50 percent more frequency yields roughly 50 percent more hashrate, provided voltage and cooling keep up. Error rate is the honest signal, because a chip producing invalid hashes still reports a high hashrate on the dashboard while contributing nothing.

Reference values

Previous-generation stock values, useful for step sizing, NOT BM1373 targets:

  BM1370  stock   525 MHz  /  1150 mV   ->  ~1.2 TH/s   @ ~15 J/TH
  BM1370  tuned   900 MHz  /  1250 mV   ->  ~1.84 TH/s  (+72%, +17.4 W)
  BM1368  stock   490 MHz
  BM1366  stock   485 MHz

  Voltage DAC range on BM-series boards:  1100 - 1300 mV
  AxeOS frequency granularity:            25 MHz steps
  Target hardware error rate:             under 2%

BM1373 stock values are not yet consistently published across vendors. Read yours from the dashboard before touching anything.

Thermal configuration

Single-chip units (Gaia, Zyber Blanc, S1-1)
  Stock fan            adequate to stock frequency only
  Recommended upgrade  Noctua NF-A6x25 (keeps unit at or below 40 dB)
  Target chip temp     below 60 C sustained
  Throttle threshold   above 70 C, drop frequency one step

Four-chip Nexus S1
  Cooling              AXP90 full-copper heatsink + heat pipe, dual fan
  Expected noise       ~50 dB
  Real power draw      budget 120-140 W on early units, not 100 W

Six-chip BSB B30
  Cooling              2 fans + heatsink, front-to-rear airflow
  Expected noise       ~50 dB
  Clearance            10 cm free at both intake and exhaust
  Room impact          400 W continuous - plan summer ventilation

Power supply

NerdAxe Gaia      PSU included, low draw
Zyber Blanc       5 V / 6 A external brick
Nexus S1          XT30 connector, PSU included by most resellers
Nexus S1-1        PSU included
BSB B30           PSU built in, 94% efficient, IEC C14
                  any standard computer power cord works

For any unit with an external supply, an undersized or electrically noisy PSU is the most common cause of instability that gets misdiagnosed as a bad chip. If you are overclocking, a quality unit such as a Mean Well LRS series buys more stability than another 25 MHz buys hashrate.

Pool and stratum configuration

Every one of these machines is a standard SHA-256 stratum client, so pointing one at a solo pool is four fields. Same hardware, same procedure on every chain. Only host and port change.

Stratum URL     stratum+tcp://<REGION>.solofury.com:<PORT>
Username        <YOUR_WALLET_ADDRESS>.<WORKER_NAME>
Password        x
Fallback pool   configure a second region as backup

Four rules that matter more than they look:

• Use an address native to the chain you are mining. A Bitcoin address in a Bitcoin Cash configuration means the pool cannot pay you. Non-custodial solo payouts go coinbase-direct to the address in your username field, so a typo is unrecoverable and no support ticket can fix it.
• Give every worker a distinct name. With more than one unit, identical worker names collapse per-device statistics into a single row and make fault diagnosis impossible. Use a scheme that survives a fleet: location, model, index.
• Always configure the fallback pool. A stratum disconnect with no fallback is silent downtime, and the dashboard keeps showing a decaying one-hour average that looks healthy for the better part of an hour.
• Pick the region closest to you. Latency does not change your block-finding probability, because the share validity window is far wider than any realistic round-trip. It does reduce stale shares at the margin.

A worker naming scheme that scales:

  Single unit         wallet.gaia01
  Mixed fleet         wallet.office-nexus01
                      wallet.office-gaia01
                      wallet.garage-b30
  Test vs production  wallet.test-zyber01   (point at eCash first)

Copy the exact host and port for your chain and region from the connection page. The full walkthrough in the solo mining setup guide applies unchanged to every BM1373 unit.

Monitoring and failure modes

Check daily     accepted vs rejected share ratio
Check daily     hardware error percentage (target: under 2%)
Check weekly    best difficulty / best share ever

SYMPTOM                          LIKELY CAUSE
hashrate 0, device powered   ->  stratum disconnect, not hardware
high hashrate, no accepts    ->  wrong address format for the chain
gradual decay over days      ->  dust or dried thermal paste
sudden drop after tuning     ->  voltage insufficient for frequency
random reboots               ->  power supply, not the ASIC
high error %, normal temps   ->  frequency above this chip's ceiling

One diagnostic worth internalising: a one-hour average hashrate above zero does not mean a miner is online. That average decays over roughly an hour after a disconnect, which is long enough to hide a fault through a whole evening. The reliable online test is time since last accepted share; anything past 10 minutes on a healthy unit means something is wrong. Our best share explainer covers what the difficulty numbers on your dashboard actually represent.

What comes after the BM1373?

Eleven years took Bitmain from the BM1385 at roughly 200 J/TH to the BM1373 at roughly 10. Straight-line extrapolation is unwise on a process node that is already at 3nm, where each further shrink costs more and delivers less than the last. The realistic near-term gains in home mining are unlikely to come from a new chip at all.

Three things are more likely to matter first. Retail chip supply, which would collapse the 60-dollar salvage price and with it the entry cost of the whole category. Firmware maturity, where a purpose-built OS tuned to one chip’s voltage and frequency curves can extract more than a generic port. And packaging, where the B30’s built-in 94 percent PSU and IEC socket are a bigger practical improvement over an XT30 pigtail than most people credit.

The chip history in full is in our ASIC chip evolution guide, and one conclusion from it is worth repeating here: the chip does not change your odds per hash. A BM1373 on a desk has exactly the same per-hash probability as a BM1373 in an industrial farm. Only the hash count differs. The network does not know how small you are.

Which one should you buy?

Key takeaways

• Five BM1373 home miners exist as of July 2026, spanning 2.4 to 30 TH/s and 26.63 to 166.66 US dollars per terahash.
• The honest generational efficiency gain over the BM1370 is about 33 percent, not the 50 percent commonly advertised. Only well-cooled single-chip and four-chip designs reach the chip’s 10 J/TH rating.
• Your break-even electricity price depends only on J/TH, not on hashrate or purchase price. At July 2026 hashprice the whole lineup falls between $0.082 and $0.149 per kWh: all positive at $0.05, most still positive at $0.10, none positive at $0.15 or above.
• The BSB B30 wins decisively on acquisition cost and loses on efficiency, firmware transparency and running cost. Below 0.09 dollars per kWh it is the obvious pick; above 0.15 it is the worst one.
• On Bitcoin all five are lottery tickets. On Bitcoin Cash a B30 averages under two years to a block, and on eCash about three weeks, though an eCash block is worth roughly ten dollars.
• The cheapest way to validate a solo configuration end to end is to point it at a low-difficulty chain first and watch a real coinbase transaction arrive.
• Tune by measured error rate in 25 MHz steps, never by copied settings. Early BM1373 silicon varies significantly unit to unit.

Specifications sourced from vendor product pages at HeliumDeploy, BSB Miners, Bitcoin Merch, XC Miner, TinyChipHub, Apexto and PunkHash, and from published bench measurements, as of 20 July 2026. Network figures: Bitcoin difficulty 127.17 T at the 11 July 2026 retarget, hashprice 31.10 dollars per PH/s per day; Bitcoin Cash difficulty 428.92 G and eCash hashrate 91.81 PH/s as of 18 July 2026. Vendor-claimed figures are labelled as such throughout. Break-even and mean-time-to-block figures are calculated from those inputs and move with the network; verify current values before making a purchase.