kHeavyHash algorithm: Kaspa, ASIC miners and specifications

kHeavyHash is a proof of work algorithm written for one specific network, Kaspa. Described in words it is simple, in substance it is unusual: a matrix multiplication wedged between two Keccak hashes. First the block header with the nonce is folded into Keccak, then the resulting numbers are run through a matrix of pseudorandom values, and the result goes back into Keccak. The comparison with the target is the same as anywhere else.

The matrix in the middle was not chosen for looks. Matrix multiplication is exactly the operation that optics handles well, and the authors were laying the groundwork for optical computing: the idea is that once such hardware exists it will deliver a jump in efficiency that ordinary silicon chips cannot reach. For now that remains a bet on the future, and the algorithm is computed by perfectly ordinary ASICs.

The practical feature of kHeavyHash is a different one, and far more important for a buyer: this is a single coin algorithm. The Kaspa forks that started on it moved to variants of their own in order to cut off ASICs, and today the entire power of the algorithm stands on Kaspa. The asic.es catalog holds 19 kHeavyHash models from four manufacturers, with efficiency spread from 116.67 to 650 J/TH, a gap of more than five times.

The work of a miner is the usual one: find a nonce that makes the final header hash land below the current target. What differs is the path that hash takes. In kHeavyHash a matrix multiplication sits between the two Keccak calls, and that is precisely what decides how an efficient chip for this algorithm has to look.

The matrix is built deterministically from the header state, so it is the same for everyone computing the current block and it changes from block to block. Preparing tables in advance does not work: every nonce is a new vector that has to be run honestly through the multiplication. Memory requirements are modest, so the silicon leans on arithmetic and on multiplier density rather than on buffer size as in Scrypt.

Two practical things follow from that. First: hashrate numbers on kHeavyHash are small, from hundredths of a terahash up to three dozen, because a single operation here is heavier than in SHA-256. Second, and this one matters more: the J/TH of this algorithm and the J/TH of Bitcoin are different units. A machine at 116 J/TH on Kaspa and a machine at 12 J/TH on Bitcoin cannot be compared with each other, neither in work nor in money.

The design of the network itself is worth holding in mind separately. Kaspa is not a chain but a blockDAG: blocks found at the same moment are not discarded, they are built into a shared graph and ordered by the GHOSTDAG protocol. That is why the network lives calmly at ten blocks per second, and the share of orphaned blocks, the usual trouble of fast chains, does not eat into miner income here.

1. Block headerVersion, parent blocks of the DAG, time, target, nonce
2. First KeccakThe header is folded into a fixed state
3. Matrix constructionA matrix of values is derived deterministically from that state
4. Matrix multiplicationThe vector from the hash runs through the matrix, the core of the algorithm
5. Second Keccak and comparisonThe result is hashed and checked against the target, otherwise a new nonce

• 2018The PHANTOM protocol and its practical version GHOSTDAG are published, the mathematics of ordering blocks in a graph instead of a chain.
• 2021Kaspa launches in November. A fair start with no premine and no presale, mining runs on graphics cards.
• 2022Smooth emission begins in May: the reward drops every month rather than all at once every four years. The kHeavyHash algorithm is designed with optical computing in view.
• 2023The first ASICs appear. IceRiver ships the KS line, and Bitmain follows with the Antminer KS3. Graphics cards leave the network within a few months.
• 2024The Antminer KS5 and KS5 Pro arrive at 150 J/TH and set the benchmark for a long while. IceRiver answers with the KS5L and KS5M.
• 2025The Crescendo fork lands in May: the network moves from one block per second to ten, the block reward is divided by ten, and emission per second stays the same.
• 2026The current ceiling of the catalog is the IceRiver KS7 at 30 TH/s and 116.67 J/TH. By mid year about 95% of the whole Kaspa emission has been mined.

Below are the twelve highest performing machines in the catalog. The full list of 19 models opens with the button, including the entry level KS0 and KS1 that people take for home use and for a first look at the algorithm.

There is no custom firmware for kHeavyHash in the catalog, and that has to be understood before buying. AsicBoost firmware is built for SHA-256 boards, on Scrypt it covers only two models, and the KS series and IceRiver are not on that list. Network installation through HashCore Toolkit belongs to the same supported lines, it is not meant for Kaspa machines.

What that changes in practice. Autotuning of each board, manual frequency and voltage profiles, careful undervolting for the sake of J/TH are all unavailable here. You work in the modes the manufacturer set, and the whole of optimisation comes down to cooling, to the quality of your power and to which model you pick when buying.

A simple conclusion follows: on kHeavyHash the efficiency of a machine is fixed at the moment of the order. On SHA-256 a poor purchase can be partly rescued with firmware and undervolting, here there is no such headroom, and the gap between 150 and 400 J/TH stays with you for the whole service life of the hardware.

A separate word about third party builds found online. They are offered for KS boards, but with no clear support and with the warranty gone. Without a mature ecosystem that is a lottery where the win is measured in percent and the loss in the price of a board.

The income of a kHeavyHash farm comes from a single coin, so the picture is simpler and harsher than on other algorithms: Kaspa network difficulty, the KAS exchange rate and your efficiency in J/TH. There are no side payouts here of the kind merge mining brings on Scrypt.

Emission works differently from Bitcoin, and that is worth counting into payback. The reward drops by roughly six percent every month, adding up to a full halving over a year. There will be no sharp step you can prepare for in advance, but there will be a constant slow slide of income at unchanged difficulty.

A second factor of the same order: by mid 2026 about 95% of the whole emission has been mined. From here the contribution of the block reward to miner income will only shrink, while the role of transaction fees grows. Counting payback three years ahead at the reward of today makes no sense, budget for the decline.

Open the profitability calculator →

Efficiency here is measured in joules per terahash, J/TH, as in SHA-256, but that is a coincidence of notation, not of units. A kHeavyHash terahash and a SHA-256 terahash are different work, so comparing 116 J/TH on Kaspa with the figures of Bitcoin machines is not valid. The comparison only makes sense within the algorithm.

Reference points from the catalog for 2026. From 116 to 150 J/TH is the current generation, the IceRiver KS7 and KS7 Lite, the Antminer KS5 and KS5 Pro. From 226 to 340 J/TH is one step back, the KS5M, KS5L, KS2 Lite, KS0 Ultra and the iBeLink BM-KS Max, which live on cheap power. From 370 to 400 J/TH are the Antminer KS3 and IceRiver KS3, hardware already looking at the exit. Above 500 J/TH there are only the entry level KS0, KS1, KS2 and KS3L, the home segment, where the point is interest rather than money.

Translated into costs in practice. A machine at 116 J/TH running 30 TH/s burns 3.5 kW, while a machine at 400 J/TH would need about 12 kW for the same 30 TH/s, that is three and a half units of 3.2 kW instead of one. A gap of 8.5 kW running around the clock is about 74,000 kilowatt hours a year, a sum that in Europe buys several more machines like it.

Settle the coin question first. You are buying hardware for one network with no fallback scenario. If that concentration of risk does not suit you, look at SHA-256 or Scrypt, where the fleet can be switched.

Count the tariff, not the hashrate. Above 0.10 euro per kilowatt hour only machines up to 150 J/TH stay in the game. At 0.05 euro and below, models of the 226 to 340 J/TH generation still make sense.

Do not count on firmware. What you buy is what you get. There is nothing to change the mode the manufacturer set, so paying up for better efficiency at the start pays off here more than on other algorithms.

A home machine is not a small farm. The KS0 at 65 watts and the KA-BOX at 400 watts are good as a quiet device in a flat. Putting them in a rack is pointless: at 339 to 650 J/TH the economics do not add up.

Check power and cooling. Industrial KS units eat 3 to 3.5 kilowatts and are as loud as any ASIC, around 75 decibels. A dedicated power line and supply and exhaust ventilation are needed from day one, not later.

Budget for a falling reward. The monthly slide of emission and the 95% already mined mean that income in KAS will fall on its own. The machine has to pay off with room to spare, not break even on the numbers of today.

Comparing J/TH with Bitcoin machines. The same unit notation does not make the work the same. Putting 116 J/TH on Kaspa next to 12 J/TH on Bitcoin is not valid, these are different algorithms and different terahashes.

Waiting for custom firmware. There is no mature ecosystem for kHeavyHash and there may never be one: the market is too narrow and there is a single coin. Plan the purchase as if firmware will never appear.

Taking an entry model as a trial and scaling later. The KS0 and KS1 at 600 to 650 J/TH do not scale: ten such machines will eat more electricity than one KS7 and deliver less hashrate.

Counting payback at the reward of today. The block reward drops every month. A calculation two years ahead at the current figure overstates the result far more than it does in Bitcoin with its rare halvings.

Ignoring the ping to the pool. Ten blocks per second forgive much less than one block every ten minutes. A pool on another continent means stale shares and a lost percentage of income.

Buying used KS units without checking the boards. Liquidity and repairability on these machines are lower than on Antminer units for Bitcoin. A dead board on a KS3 often means the machine is easier to strip for parts than to fix.