Bitaxe Duo 650: A Density Play, Not a Value Play

The Bitaxe Duo 650 puts two BM1370 chips in the same footprint as a single-chip Gamma, for $104. It was the best value per terahash in the range when it launched. Then the Gamma dropped to $61 and that stopped being true. What the Duo still offers is density, and one thing the spec sheet never mentions that thirty-two owner reviews agree on.

This is a review built from two sources: the published design and manufacturer figures, and what owners actually measured. Where those two disagree, we say so.

Key takeaways

• $63.98 per TH/s at stock, against $50.83 for a Gamma at current prices. The Duo is no longer the cheapest hashrate in the lineup, and its own seller’s “best value per terahash” claim predates the Gamma price cut.
• 1.63 TH/s at 25.8 W, about 15.8 J/TH, with a stated tolerance of plus or minus 10 percent.
• Owners report 2.1 to 2.67 TH/s overclocked, and one reports 13.5 J/TH at 2.4 TH/s — better efficiency than stock, which points to undervolting alongside the frequency bump.
• Replace the power supply. At least six of the thirty-two reviews say so independently. One owner reports losing two boards to power events.
• 36% more hashrate per device in an identical 4 by 2.25 by 2 inch footprint, so existing stands, enclosures and 5V supplies carry over. That is the actual case for buying one.
• The BM1370 is a 5nm chip, not 3nm. Some retail listings say otherwise; Bitmain’s first 3nm SHA-256 part is the BM1373.

What it is

The Duo 650 is the 65x member of the Bitaxe family: two BM1370 ASICs on a board that keeps the exact dimensions of the single-chip Gamma. Design files are published at bitaxeorg/bitaxeGamma on the gamma-duo-650 release tag, under the same open licence as the rest of the line.

Manufacturer figures, read 7 August 2026, with a stated tolerance of plus or minus 10 percent.

One correction worth making. At least one retail listing describes the BM1370 as a 3nm chip. It is not: the BM1370 is a 5nm part from the S21 Pro generation, and Bitmain’s first 3nm SHA-256 silicon is the BM1373 that ships in the S23 series. The distinction matters because 3nm buys a real efficiency jump — roughly 10 J/TH against 15 — and it is not what you are getting here. Our ASIC chip evolution guide traces the whole lineage, and the BM1373 era covers what actually changed at 3nm.

Cost per terahash: the argument that expired

When the Duo launched it was the cheapest hashrate in the Bitaxe lineup, and its seller still says so. That claim was written when a Gamma cost $97.98. The Gamma now lists from $61.

Prices are the low end of each live listing on 7 August 2026, boards only, excluding power supplies and stands. Efficiency figures are manufacturer or bench values, not wall measurements.

The Duo now sits fourth. A Gamma delivers hashrate 26% cheaper, and even a NerdQAxe++ at four times the price beats it per terahash. If your only criterion is odds per dollar, buy Gammas.

The comparison against two Gammas makes the same point from the other direction. Two Gamma boards cost $122 and give you 2.4 TH/s. One Duo costs $104.29 and gives you 1.63 TH/s. For $17.71 more you get 47% more hashrate — but you also get two devices, two power supplies, two Wi-Fi connections and two spots on the desk.

That trade is the whole product. The Duo is not the cheapest hashrate; it is the most hashrate you can get from one of everything.

One genuine spec win survives all of this: at 15.8 J/TH the Duo is the most efficient board in the single-board range, marginally ahead of the Gamma’s own bench figure of 16.6 J/TH at stock. Two chips sharing one regulator and one board is simply a more efficient arrangement than two separate boards, and over years of continuous running that difference is real if small.

A note on prices generally. This market moves fast in one direction: a Gamma listed near $98 in February 2026 and near $61 in August. Any cost-per-terahash table, including this one, is a snapshot. Check the live listings before deciding.

What owners actually measured

The product carries thirty-two published reviews averaging 4.88 out of 5. Read as a dataset rather than as testimonials, they are more informative than the specification.

Owner reports published February to August 2026. Individual results, not a controlled test.

Two things stand out.

The 13.5 J/TH result is better than stock. That is not what a plain overclock does, since raising frequency alone costs efficiency. Getting more hashrate and better efficiency means the owner undervolted while raising frequency, which is the tuning approach that actually pays on this silicon. It also tells you the stock settings leave headroom on the table.

Measured draw runs above spec. One owner reports 27.1 W against the stated 25.8 W. That is a 5 percent gap at the board, and the gap at the wall socket is larger still, because the 5V-to-core conversion and the ESP32 both consume power the dashboard does not count. If you are budgeting electricity, add 15 to 20 percent to any dashboard reading.

The power supply is the real story

Nothing in the specification tells you this. Six separate reviewers do.

The recommendation to replace the stock adapter with a Mean Well unit appears independently across the review set, in the plainest terms: skip the stock supply, put a Mean Well in the basket at the same time, save the money on the bundled one. Two reviewers describe the stock adapter running hot, with one adding heatsinks to it and reporting the temperature dropped substantially. One reports losing two boards to power outages while other Bitaxe units on the same surge protector survived.

The mechanism is not mysterious. Two BM1370 chips on a 5 V rail pull real current, and 5 V systems are unforgiving of voltage drop: the same wattage at 5 V means five times the current of a 25 V system, so every milliohm in the adapter, the cable and the barrel plug turns into heat and sag. One reviewer makes exactly this point, noting that a good quality barrel jack matters as much as the adapter and that high resistance shows up as both voltage drop and temperature.

The practical version:

1. Minimum is 5 V at 6 A, or 30 W, on a 5.5 by 2.1 mm centre-positive barrel jack. Never 12 V — it will destroy the board.
2. If you intend to overclock, go higher. Owners reporting the best results used 8 A and 100 W supplies.
3. The cable and plug matter. Thin conductors and cheap barrel plugs cause the voltage drop that makes a board look unstable when the adapter is fine.
4. Put it on surge protection, and treat power events as a real failure mode rather than a theoretical one.

If your board is already behaving oddly — high error rates, unexplained restarts, hashrate below nameplate — power is the first thing to rule out, and our troubleshooting guide walks the full diagnosis in order.

The odds, honestly

Calculated as difficulty multiplied by 2³² divided by hashrate, at BTC difficulty 126.23 trillion and BCH difficulty 430.03 billion, both read in early August 2026. Difficulty moves; check the live Network Radar before planning around any figure here.

On Bitcoin this is a lottery ticket, and the daily odds at stock are roughly one in 3.85 million. What the Duo buys you over a single Gamma is a 48 percent improvement on a number that stays astronomically small either way. Our documented Bitaxe wins page covers the blocks that have actually landed, including a Gamma at 995.2 GH/s taking a full block in July 2026.

The other chains are where a device this size changes character. Thirty-six years on Bitcoin Cash is still long, but it is a number rather than a metaphor, and the smaller SHA-256 chains bring it down much further. Our best coins to solo mine page has the ladder, with the honest caveat that easier blocks are worth proportionally less.

Where it falls short

The most useful review in the set is the three-star one, and the manufacturer’s published reply to it is worth reading alongside.

The complaint: the reviewer’s tuned Gamma boards reach higher hashrate than the Duo does at stock, getting to 2.0–2.1 TH/s on the Duo requires pushing it hard, two ASICs on a small board generate concentrated heat, and cooling options within that footprint are limited without custom metalwork.

The manufacturer’s answer, publicly posted: the Duo was designed as a budget board that performs well out of the box without upgrades, positioned between the Gamma and the GT, and it was never intended for heavy overclocking. The regulator is rated to 100 °C and production units run well inside that.

Both are right, and together they define the product. The Duo is a density board meant to run near stock. It is the wrong board if tuning is the hobby. For that, the GT 801 platform has the beefier power stage, the 60 mm fan and the 12 V input that make aggressive overclocking sensible.

Three more limits worth knowing. Wi-Fi is 2.4 GHz only, which surprises people on 5 GHz-only networks. Noise at 40–50 dB is higher than a single-chip Gamma, because two chips need more airflow in the same volume. And the warranty is 90 days.

Who should buy one

1. Anyone adding a board where space or outlets are tight. Same footprint, same stands, same 5V supplies, same firmware, 36% more hashrate than another Gamma in the same physical slot. The upgrade path is frictionless even if it is no longer the cheapest.
2. First-time buyers who want plug-and-play. Reviewers repeatedly describe five-minute setup and rock-solid running at stock. If you do not want tuning to be part of the hobby, this is the board that does not ask you to.
3. Anyone constrained by sockets, desk space or Wi-Fi slots rather than by money. If the limit is how many devices you can physically run, the Duo gives you 36% more hashrate per slot than a Gamma.

Buy a Gamma instead if odds per dollar is your only criterion — it is 26% cheaper per terahash and you can simply buy two. Buy a GT if you enjoy aggressive tuning, or something else entirely if you need the quietest possible device in a bedroom.

And be clear-eyed about the economics. At 1.63 TH/s the expected Bitcoin return is a fraction of a cent per day against $2 to $7 a month in electricity. That is not an income calculation, it is the price of a ticket. What makes it a fair ticket is that it is verifiable and non-custodial: if the block lands, the coinbase pays your address directly, with nothing to claim and nobody holding it.

One last note on where you buy. Because the design is open, anyone may manufacture a Duo, and quality varies. Before ordering from an unfamiliar seller, our guide to spotting a bad clone covers the board-version check, the ESP32 module part number and the firmware string that separate a good replica from one that will cost you.