New algorithms · ASIC design

An ASIC for a new proof of work

Launching a coin, or asking whether your algorithm can have an ASIC? We design mining ASICs for new proofs of work, and accelerators for zero-knowledge proofs. Every proof of work is compute, and compute can be accelerated: we work out with you how much of it to fix in logic, where the memory should live, and what the chip earns against a GPU, before a mask set.

New proofs of work

Can our proof of work have an ASIC?

Yes, and the useful question is how much it gains. A hash pipeline of fixed logic gains the most; a proof of work that leans on memory bandwidth or random programs gains less, and its accelerator looks more like a processor. A proof of work built on matrix multiplication sits between them: it runs on the multiply-accumulate arrays that AI accelerators are built from. Kaspa's kHeavyHash, a 64 × 64 matrix of 4-bit values between two hashes, is one example.

We design
Mining ASICs, hashboards and miners for new proofs of work; accelerators for zero-knowledge proofs
Before a mask set
Energy per hash, performance and return weighed against CPUs, GPUs and FPGAs
Matrix arithmetic
The multiply-accumulate arrays of our NPU design work
Our record
Six mining chips in volume production, five of them at customers
Algorithms our mining work spans
7
Mining chips in volume production, five of them at customers
6
Chip-level design figure of our 7 nm SHA-256 ASIC, at 0.34 V
25 J/TH

The program

From your algorithm to tested machines

  1. The algorithm

    Your specification, a reference implementation and test vectors, under a mutual NDA.

  2. Feasibility and economics

    Which steps can be fixed in logic, which must stay programmable, how much memory each needs, and what the chip earns per unit of energy against the best CPU, GPU and FPGA, worked out before a mask set.

  3. Architecture

    A hash pipeline, a multiply-accumulate array, a many-core processor or a mix, sized to the algorithm and to the node.

  4. RTL and FPGA verification

    The design in RTL, checked against your test vectors and verified on FPGA before layout.

  5. Full-custom design

    Custom cells, hand placement and libraries re-characterised for near-threshold voltage, where they pay off in energy per hash.

  6. GDS out

    Signed-off GDS for production silicon, as for the five mining chips we designed for customers that entered volume production there.

  7. Hashboards and firmware

    The hashboard and firmware are our own design; the control board, power supply and cooling are developed jointly with leading industry partners.

  8. Prototype machines and acceptance

    Prototype machines of our design, built with our partners and run against an acceptance test agreed in the contract, on your pool or network.

  9. Production

    Built with our EMS partners; more than 100,000 miners have been built on our chips and designs.

Before you sign

The figures to ask any vendor for

Ask every vendor for the same figures, measured the same way. These are the ones we put in a proposal.

The figureWhat we give youWhy it matters
Hash rate per chipAn estimate with its basis: the clock, the voltage and the pipeline it assumes.Sets how many chips a machine needs, and what each wafer is worth.
Hash rate per machineStated per machine, at the voltage and clock of each operating mode.What the site is paid for: chips per hashboard, hashboards per machine, at a stated voltage.
Power at the wallEnergy per unit of hash rate at the socket, with cooling included.Chip-level efficiency leaves out the power supply, fans and control board; the site pays for all of them.
Engineering cost apart from hardwareEngineering quoted apart from chips and machines, stage by stage.Engineering is paid once and hardware per unit; quoted apart, each can be checked on its own.
Minimum orderStated in the proposal, with the lot sizes behind it.Masks and wafers come in fixed lots, so a vendor's minimum follows from them.
Acceptance testA test written into the contract: hash rate and power at the wall, on your pool, over an agreed period.A written test lets each delivered machine be accepted on measured figures.
ScheduleA date for each stage, with what it depends on.First silicon, prototype machines and production each have their own date.

Where AI chips and mining chips meet

Matrix arithmetic, from NPUs to proofs of work

  1. Multiply-accumulate arrays

    The compute core of an NPU and of a matrix proof of work is the same: an array of small multipliers and adders, where full‑custom logic pays in energy per operation.

  2. Our NPU design work

    Our full‑custom plans for NPUs cover MAC array customisation, evaluation against a baseline, block-level sign-off and compute-core test, and we have carried out 6 nm AI full‑custom chip design for a top AI customer.

  3. Loaded once, used for every nonce

    Where the matrix depends on the block and not on the nonce, as in kHeavyHash, the chip loads it once per job and spends its silicon on the multiply and the hashes.

  4. Low-precision integers

    Matrix proofs of work use small integers, not floating point, so each multiplier is a small block that can be laid out by hand and repeated across the array.

  5. Near-threshold voltage

    The methods that let our 16 nm SHA-256 ASIC hash at a core voltage as low as 0.38 V apply to any datapath that repeats one block many times.

  6. Test and binning

    Our patent on test and binning of many-core chips sorts a large array into grades, so each chip runs at the voltage it suits.

For a first answer

What we need from you

  1. The specification

    The algorithm as written for implementers: every step, every constant, and what may change in a later fork.

  2. Reference code and test vectors

    A reference implementation and known inputs and outputs, so the design can be checked from its first line.

  3. The network figures

    Block time, difficulty adjustment and the hash rate the network runs today, or is expected to.

  4. Your target

    Hash rate per machine or per site, the power per machine and the price of power where it will run.

  5. Volume and date

    How many machines, and when they must hash.

  6. How you will accept

    The pool or network test you will accept the machines on.

Questions about ASICs for new algorithms

Can an ASIC be built for any proof of work?

Yes, at the right balance. Every proof of work is compute, and compute can be accelerated; resistance to ASICs raises the cost of the accelerator and changes its design. We weigh fixed logic against programmable logic and memory on the die against memory beside it, before a mask set.

Our proof of work uses matrix multiplication. Does that suit an ASIC?

Yes. Matrix multiplication of small integers runs on multiply-accumulate arrays, the core of every NPU, and our full‑custom techniques serve those arrays in AI chips and in mining chips alike.

How do you estimate hash rate before there is silicon?

From the architecture: the pipeline's operations per hash, the clock the node and voltage allow, and the number of cores per die, checked on FPGA before layout. The estimate comes with its basis, and the first silicon is measured against it.

Is the engineering cost separate from the hardware?

Yes. Engineering is quoted apart from chips and machines, stage by stage, so you can see what you pay once and what you pay per machine.

How are machines accepted?

Against a test written into the contract: hash rate and power at the wall, on your pool or network, over an agreed period.

What are the payment terms?

Agreed per contract.

Do you design the whole machine?

We design the chip, the hashboard and the firmware; the control board, power supply and cooling are developed jointly with leading industry partners, and the machines are built with our EMS partners.

Do you design zero-knowledge proof accelerators?

Yes, alongside mining ASICs for new proofs of work.

Ask about your algorithm

Send us the specification, the reference code and your target. We will reply with how much an ASIC gains and the stages to get there.

Coin and algorithm facts are from each coin's own documentation and public records, and our figures are our own chip-level measurements, the specifications of our designs or our own records.