Blake3 algorithm: the Alephium network, ASIC miners and specs

Blake3 is a cryptographic hash function published in January 2020. The authors, Jack O’Connor, Jean-Philippe Aumasson, Samuel Neves and Zooko Wilcox-O’Hearn, took the proven BLAKE2 as their base and added the Bao tree mode to it. Inside, nothing is computed as one long chain; it is a Merkle tree. The input is cut into 1024 byte chunks, every chunk is hashed independently, and the results are then folded upward in pairs to the root.

That tree makes Blake3 exceptionally fast and almost unlimited in parallelism. For cryptography that is a virtue, while for proof of work it would seem to be a flaw: PoW algorithms usually try to be expensive on memory or slow instead, in order to delay the arrival of dedicated chips. The Alephium network, the only one where Blake3 is mined on ASICs, did exactly the opposite and stated it in plain words: it does not resist ASICs, it is designed for them.

The asic.es catalog holds 6 Blake3 models from three manufacturers. The efficiency spread here is the narrowest of all our algorithm pages: from 225 J/TH on the Antminer AL1 to 500 J/TH on the Goldshell AL-BOX, a little more than twofold. The spread in market prices, on the other hand, is enormous, and that gets its own section below.

The miner task is the usual one: pick a nonce so that the block header hash lands below the current target. What is unusual is the hash itself. Blake3 works as a tree: the data is cut into 1024 byte chunks, inside a chunk it runs in 64 byte blocks, and the compression core does seven rounds instead of the ten of its predecessor. The leaves of the tree are computed independently of each other and are then folded in pairs down to a single root value.

It is precisely the independence of the leaves that makes the function so fast. On an ordinary processor it beats BLAKE2 several times over thanks to vector instructions, and in silicon it becomes an ideal candidate: the circuit can be repeated hundreds of times on one die and run all at once. No memory, no random access, no branching to get in the way of packing them densely on the die. Hence the thousandfold gap between a graphics card and an ASIC.

In Alephium the header is passed through Blake3 twice, the way Bitcoin does it with SHA-256. There are two quirks. First: the nonce is 24 bytes long and sits before the header rather than after it. Second: the chain number is encoded into the block hash, so a solution that is found belongs to one specific shard and cannot be reused in another. The node hands the pool sixteen templates at once, one per chain.

On top of this the network has a Proof of Less Work mechanism built in. Once total hashrate reaches 1 EH/s, miners will start burning most of the reward above that threshold, meaning that growth in network power will run into economics rather than into hash difficulty. As of July 2026 hashrate holds in the single digit petahashes, the threshold is far away and the mechanism is not switched on.

1. Block headerVersion, links to the parent blocks, roots, time, target
2. A 24 byte noncePlaced before the header, whereas the usual practice is to put it after
3. First Blake3The data is cut into 1024 byte chunks and the tree is folded to the root
4. Second Blake3The result of the first pass is hashed once more
5. Comparison with the targetBelow the target and the block is found. Above it, a new nonce and all over again

• 2020In January Blake3 is published at the Real World Crypto conference. The base is taken from BLAKE2, the tree mode from the Bao project.
• 2021On 8 November the Alephium mainnet launches. Blake3 is chosen deliberately as an ASIC friendly algorithm rather than an ASIC resistant one.
• 2024In February the ALPH rate hits its all time high. The network is considered entirely graphics card driven, there are no machines for it yet.
• 2024In May the Goldshell AL-BOX gets the ball rolling: the first production ASIC for Blake3.
• 2024In June the Rhône upgrade cuts block time from 64 to 16 seconds.
• 2024In July Bitmain ships the Antminer AL1, then the AL1 Pro. In September IceRiver answers with the AL line. The graphics card era ends within a few months.
• 2024In November the Goldshell AL-MAX closes the wave of releases. After it nobody has shown a new chip for the algorithm.
• 2025On 15 July comes the Danube upgrade: block time falls to 8 seconds, and the fixed supply cap is removed and replaced with an endless tail issuance. Reference sites still write about a billion coins.

All 6 Blake3 models in the catalog, sorted by efficiency with the best on top. Hashrate in TH/s, efficiency in joules per terahash. Click a name to open the profitability calculator for that model.

Blake3 is the only algorithm in this series where third party overclocking firmware exists for the catalog machines, but it is too early to celebrate. AsicBoost firmware is built for SHA-256 boards, on Scrypt it covers two models, and there are no Alephium machines on that list. Network installation through HashCore Toolkit belongs to those same supported lines.

What the market does have. For IceRiver machines an enthusiast has released an open firmware with overclocking: a frequency shift across a wide range, fine voltage tuning, applying changes without a reboot, a smooth frequency ramp in small steps. The hashrate gain is roughly proportional to the frequency gain. The developer fee is one percent, or 1.33 percent in the hosting build.

Three reasons to weigh it all up twice. First: the author himself warns that there are no software limits on frequency and voltage, so it is easy to bring the machine down or to overheat the power stage. Second: the last version came out in October 2024 and there have been no updates since. Third: the warranty ends after such a flash. If you decide to go ahead, start with small steps and watch the temperatures of the power stages, not only of the chips.

For the Antminer AL1 and the Goldshell boxes we did not find a single custom firmware. For those machines the factory efficiency numbers are your numbers for the whole time you own them.

Let us start with the unpleasant part, because that is the honest way. By July 2026 the ALPH exchange rate sits at roughly one percent of the February 2024 all time high. At the electricity prices normal for Spain and the EU, not one machine on this algorithm comes out in the black. Buying Blake3 hardware today makes sense in two cases only: your electricity is very cheap or your own, or you are deliberately betting on the project and are ready for payback never to arrive at all.

Formally income is set by three quantities: network difficulty, the ALPH exchange rate and your efficiency in J/TH. But Alephium has a feature that none of the other algorithms in this series has. The block reward is taken as the lower of two curves: one depends on time, the other on network hashrate. There is no classic halving, the decline is continuous and is recomputed every block. All transaction fees are burned in full.

A separate word about the price of the hardware itself. The IceRiver AL3 gives practically the same hashrate as the Bitmain flagship and costs several times less on the secondary market. If you do go into this algorithm, compare not only J/TH but also the price per terahash: the gap there is bigger than the gap in efficiency.

Open the profitability calculator →

Efficiency here is measured in joules per terahash, J/TH. The notation matches SHA-256, Eaglesong and kHeavyHash, but that is a coincidence of notation, not of units. A Blake3 terahash and a SHA-256 terahash are different work, so putting 225 J/TH on Alephium next to the single digit figures of Bitcoin machines is meaningless. The comparison only makes sense within the algorithm.

Inside it the spread is small by the standards of other pages. The Antminer AL1 needs 225 J/TH, the IceRiver AL3 needs 233.33, the IceRiver AL0 needs 250, and the three Goldshell boxes hold at 480 and 500. Between the leader and the bottom row it is a little more than twofold. That means the difference in the electricity bill on Blake3 is not as dramatic as on other algorithms, and the price of the machine itself comes to the front.

The arithmetic goes as usual: multiply efficiency by hashrate to get draw in watts, divide by 1000 to get kilowatts, multiply by your tariff and by 720 hours a month. Compare the sum you get with the mining forecast in the calculator. At the current rate the result will most likely not please you, and it is better to see that before the purchase.

Do the numbers first, decide afterwards. This is the only algorithm in our series where we say outright: start with the calculator and your real tariff. At an ordinary electricity price the answer will be negative, and that is a normal result of the calculation, not an error.

Compare the price per terahash, not just J/TH. The efficiency difference between the flagships is under five percent, while the price difference on the market is a multiple. Here the hardware is picked with money, not with watts.

Settle the coin question in advance. A Blake3 machine has no fallback network. Decred on the same hash function runs under a different scheme and is not mined with this hardware. You are buying a bet on one project.

Keep the home scenario separate from the rack one. The IceRiver AL0 at 100 watts and the Goldshell boxes at 180 to 480 watts live in a flat. The Antminer AL1 and the IceRiver AL3 need about 3.5 kilowatts, their own breaker and an exhaust setup.

Do not rush into firmware. Overclocking is available only for IceRiver, from a third party developer, with no software limits and no updates since the autumn of 2024. It voids the warranty. On a new machine it is better to work through the stock mode first.

Do not build a rate recovery into the calculation. Forecasts of ALPH climbing back to its former levels are hope, not arithmetic. Run the numbers as they stand today, and let growth, if it comes, be a pleasant surprise.

Assuming a Blake3 machine will take Decred. The same hash function does not mean the same proof of work. Decred has a single Blake3 over a header of a different length, Alephium a double one with the nonce in front and the chain number inside. The hardware is not interchangeable.

Taking supply figures from old reference sites. The cap of one billion ALPH was removed by the Danube upgrade in July 2025 and replaced with an endless tail issuance. Many sites still have not reflected that.

Comparing J/TH with Bitcoin machines. The same unit notation does not make the work the same. Putting 225 J/TH on Alephium next to the single digit figures on Bitcoin is not valid.

Buying on profitability numbers from 2024 advertising. Promises of hundreds of dollars a day belonged to the period before mass hardware arrived. The picture today differs by orders of magnitude, and it takes a minute to check in the calculator.

Flashing firmware straight out of the box. Third party overclocking with no limits on a machine that has not been run in is the fastest way to a dead board and a voided warranty. Stock mode and measurements first.

Chasing hashrate and forgetting the price. On this algorithm two machines four percent apart in efficiency can differ several times over in price. Work out the cost per terahash.