Skip to content
qubitwerkQuantum simulation by Coolblack
Home

Our research. Results, not promises.

What we have measured, which programs it applies to and what comes next. Results are public; our internal implementation remains private.

Selected performance examples · 21 October 2026

Faster calculations on our PC.

We compared our simulator with Qiskit Aer, an established reference simulator, on the same i9 with RTX 3090. Our simulator took less time in these four selected examples.

You see recorded timings and calculated results. Timings apply to follow-up requests in a running process, for these circuits and readout queries.

Checked against Aer

22 qubits on our graphics card

22 logical qubits

25.558 msOur simulator · median follow-up time
Aer GPU: 41.294 ms1.62 × as fast in this test
CALCULATED RESULTZ₀ = -0.203986

In the model, qubit 0 has about 60.20% probability of 1 and 39.80% of 0. The Z value is the probability of 0 minus the probability of 1.

Measurement conditions and accuracy
Timing
Median of four follow-up requests per simulator across two separate processes. Timings cover simulator computation requests, not isolated GPU kernel time.
Local response time, including transport
Our simulator: 28.226 ms · Aer GPU: 42.557 ms. Excludes the public queue, internet and animation.
Checked results
64 selected amplitudes, norm and 22 Z expectation values. We did not compare the full state for this example.
Largest checked amplitude error
2.638e-18 · Limit: 1e-12.
Checked against Aer

Find the treasure among 256 possibilities

8 search qubits + 1 auxiliary qubit

65.794 msOur simulator · median follow-up time
Aer GPU: 99.382 ms1.51 × as fast in this test
CALCULATED RESULT11111111 · 99.9947 %

This is the calculated chance of finding the specified target after this search program. A single measurement remains random.

Measurement conditions and accuracy
Timing
Median of four follow-up requests per simulator across two separate processes. Timings cover simulator computation requests, not isolated GPU kernel time.
Local response time, including transport
Our simulator: 75.994 ms · Aer GPU: 109.482 ms. Excludes the public queue, internet and animation.
Checked results
We checked the full state of 512 amplitudes, norm and 9 Z expectation values, also against the original circuit. The full-state check is separate from the timings above.
Largest checked amplitude error
9.045e-15 · Limit: 1e-12.
Checked against Aer

Find the treasure among 64 possibilities

6 search qubits

17.995 msOur simulator · median follow-up time
Aer GPU: 43.072 ms2.39 × as fast in this test
CALCULATED RESULT111111 · 99.6586 %

This is the calculated chance of finding the specified target after this search program. A single measurement remains random.

Measurement conditions and accuracy
Timing
Median of four follow-up requests per simulator across two separate processes. Timings cover simulator computation requests, not isolated GPU kernel time.
Local response time, including transport
Our simulator: 22.89 ms · Aer GPU: 47.383 ms. Excludes the public queue, internet and animation.
Checked results
We checked the full state of 64 amplitudes, norm and 6 Z expectation values, also against the original circuit. The full-state check is separate from the timings above.
Largest checked amplitude error
3.353e-15 · Limit: 1e-12.
Checked against Aer

Two qubits, one shared state

2 logical qubits · 3 qubits simulated in the comparison

0.44 msOur simulator · median follow-up time
Aer GPU: 9.202 ms20.93 × as fast in this test
CALCULATED RESULT00: 50 % · 11: 50 %

The two results match: 00 or 11, each with 50% probability. In the calculated model, 01 and 10 have zero probability.

Measurement conditions and accuracy
Timing
Median of four follow-up requests per simulator across two separate processes. Timings cover simulator computation requests, not isolated GPU kernel time.
Local response time, including transport
Our simulator: 1.078 ms · Aer GPU: 10.685 ms. Excludes the public queue, internet and animation.
Checked results
We checked the full state of 8 amplitudes, norm and 3 Z expectation values, also against the original circuit. The full-state check is separate from the timings above.
Additional simulated qubit
Both simulators used three qubits for the comparison. The third stayed in state 0 and is not part of the entangled pair.
Largest checked amplitude error
1.110e-16 · Limit: 1e-12.
Shared measurement conditions and scope

We measured on 21 October 2026 on the Intel Core i9-10900K with NVIDIA RTX 3090. Both simulators used double precision (complex128). Reference: Qiskit Aer 0.18.0 with Qiskit 2.6.0, GPU statevector with cuStateVec.

We selected these four performance examples after the measurement series. Four follow-up values per simulator and example form a small sample. Both simulators calculated the same readout outputs. We recomputed each follow-up request; timings are not cold-start or website measurements.

This selection demonstrates advantages for these workloads under these conditions. It does not establish a general speed advantage or quantum advantage. Our PC simulates quantum programs on classical hardware. We keep our internal implementation private.

Research status · 27 September 2026

Our results on the i9

Explore available experiments, completed tests and our next goals. We name the task alongside each qubit count: a search and a particle model place different demands on the computer.

  1. Verified research result · 20 September 2026

    35 qubits simulated on our PC

    On 20 September 2026, we simulated the full state of a 35-qubit circuit: 34,359,738,368 complex amplitudes in double precision. The RTX 3090 processed the 512 GiB state from SSD in blocks. Our computer took 10.8 hours. Qiskit Aer reached 31 qubits in our in-memory comparison at the time. This capacity comparison does not establish a speed advantage.

    How we checked it →
  2. New · 27 September 2026

    A diamond as a qubit.

    At a nitrogen-vacancy centre in diamond, researchers can use two spin states as a qubit. In the diamond lab, you control a model of these states with microwave pulses and simulate 1,000 measurements. Your browser computes the model. We checked it against an independent Lindblad calculation; the largest deviation was 0.0000000046.

    Open the diamond lab →
  3. Try it now

    24 search qubits

    Set a code with six or seven digits. For seven digits, the RTX 3090 simulates the search across 10,000,000 codes and the i9 checks the result. The i9 CPU handles six digits.

    Search for your test code →
  4. In research testing

    Comparing software on 780 states

    This model distributes two excitations across 40 sites, requiring 780 states. In the test, our code took 21.15 milliseconds and the established comparison software took 191.46 milliseconds, with the same accuracy check. The speed advantage applies to this test case. The full 35-qubit circuit is a different task.

    Measurements and conditions →
  5. Another example · planned

    A molecule in the simulator

    We plan a lab for a model of two hydrogen atoms. You will be able to change their distance and see how the calculated energy changes.

    The planned molecule example →
How large can the challenge get?

Choose the search size

Choose a lock size. We calculate the search space, but only show a runtime when it has been measured.

1,000,000possible codes · at least 20 search qubitsSix-digit learning lock available

Measured: median of 0.16 s across eight integration tests, 0.08 to 0.26 s depending on the run, on all ten cores of the i9. Compute time includes process start-up but excludes queueing, transfer and animation.

Try it yourself
How did we measure this?

The comparison software is called ffsim (version 0.0.83). It also specialises in this type of quantum calculation. We tested two ways of calling it and used the faster measured option as our reference.

The current research series was conducted entirely on our Intel Core i9-10900K with 62 GiB RAM. One CPU thread per method; the RTX 3090 was not used in this series. Model: an open chain with a fixed excitation count and nearest-neighbour interactions. Not a general-purpose 40-qubit simulator.

Median of nine runs per test case and method · milliseconds
Qubits / excitationsOur simulatorEstablished comparison software
12 / 25.67 ms13.32 ms
12 / 68.4 ms34.46 ms
18 / 38.6 ms34.38 ms
22 / 311.34 ms57.45 ms
30 / 213.42 ms95.66 ms
40 / 221.15 ms191.46 ms

We compare two programs, not two computers. Our simulator and the established comparison software solve the same task on the same i9. In the specialised 40-qubit model, our program takes about one ninth of the time. Across six test cases, the ratio ranges from 2.35 to 9.05. This applies to these tests, not automatically to every quantum calculation. It does not establish a quantum advantage.

All methods received identical programs on the i9. We checked the complete model results in double precision: three program seeds, three repetitions each. Error limit: L2 below 10⁻⁹. The largest deviation of non-MPS methods from ffsim was about 3.61 × 10⁻¹⁵.

Timing starts after library imports. It includes preparation, computation, result evaluation and output serialisation; for our engine, process startup is included too. The website, network and animation are excluded. Reference: ffsim 0.0.83. Test sizes and measurement conditions are public; our implementation stays private.

Completed on the i9: 312 runs without execution failures and 68 passed functional tests. 45 Aer-MPS results missed our accuracy limit and are excluded from equal-accuracy speed comparisons. This is an internal evaluation; independent external confirmation is still pending.

Download research results as data · Measurement evidence for the six- and seven-digit locks · Project and further results

Newly computed · 20 September 2026 · on our own PC

35 qubits on a single PC.

For this 35-qubit circuit, we calculated the full state with 34,359,738,368 complex amplitudes. We used double precision without pruning amplitudes. The state requires 512 GiB. We stored it on SSD and processed the blocks on the graphics card.

Standard tool on the same computer

31 qubits

2,147,483,648 amplitudes in the in-memory comparison at the time.

Our way: the state on the SSD

35 qubits

34,359,738,368 amplitudes, 16 times as many.

Insufficient memory to run circuit … Required memory: 524288M, max memory: 64187M (Host) + 24125M (GPU)
In the comparison test on the same computer and circuit, Qiskit Aer reported this memory requirement: 524,288 MiB needed, 64,187 MiB RAM and 24,125 MiB GPU memory available. Our block-based approach also uses the SSD.

The probability distribution of our result

All 34,359,738,368 probabilities are included here. The bars come from our calculation. The line shows the theoretical distribution for large random quantum states. Our statistic is 1.9999973; the theoretical reference value is effectively 2.

0Probability divided by the average6

Our five pre-specified checks

  • At 30 qubits our result agrees with Qiskit Aer: largest difference 4,5 · 10⁻¹⁹, the limit was 1,0 · 10⁻¹⁰.
  • At 30 qubits, the inverse returns to the initial state up to floating-point error: deviation 1,6 · 10⁻¹⁴, limit 1,0 · 10⁻⁹.
  • At 35 qubits the probabilities still add up to 1: deviation 1,9 · 10⁻¹⁴.
  • At 35 qubits the distribution matches the theory: 1.9999973 instead of 2.
  • At 35 qubits, the inverse returns to the initial state up to floating-point error: deviation 1,9 · 10⁻¹⁴; each of the other 34,359,738,368 amplitudes is at most 3,9 · 10⁻¹⁷.
Computational method and comparison context

We simulated a random circuit with 1,025 gates over 20 cycles on a 5 by 7 qubit grid, using the gate pattern from Google's random-circuit experiments. We alternated between two 512 GiB state files on SSD. The RTX 3090 processed 4 GiB blocks. After each block, we saved a checkpoint from which the calculation could resume after a pause.

Cost: 10.8 hours of computation, split into 458 short jobs to leave gaps for the public queue. The run wrote 4.4 TiB. The SSD stayed below 42 degrees Celsius. Hardware: Intel Core i9-10900K, 62 GB RAM, NVIDIA RTX 3090 (24 GB), 2-TB-SSD.

Other research teams have published SSD-backed simulations. Hsu and colleagues describe 39 qubits using 256 gigabytes of RAM and 16 terabytes of SSD storage (Connection Science, 2024). Our 35-qubit run used about 62 gigabytes of RAM and a 2-terabyte SSD. We document the full circuit this PC can handle under the stated checks. We do not claim a world record.

We simulate the circuit on classical hardware. Quantum advantage would require a quantum computer to outperform strong classical methods on a comparable task with comparable output quality. Our SSD capacity does not determine this. For our full state vector, each additional qubit doubles the memory required.

The measurements as data. The state files, each 512 GiB, were intermediate data and were deleted after evaluation.

Hands-on research · recorded simulation

Explore how two excitations interact

Two excitations start on neighbouring sites. Play the recordings and compare three interaction strengths. The bars show where a measurement could find the excitations.

?
Your prediction: which interaction gives them the highest chance of being neighbours at the end?
Recording · not a new server calculationStep 0 / 32

Top: the distribution now. Below: its development from top to bottom. More intense means more likely — not a particle trajectory.

01

None

Interaction · 0
100%0%0321Site40
Chance of being neighbours100 %
02

Medium

Interaction · 0.8
100%0%0321Site40
Chance of being neighbours100 %
03

Strong

Interaction · 2
100%0%0321Site40
Chance of being neighbours100 %
Colour intensity0 %25 %50 %75 %100 %

All three views show the same program step and the same scale. There are always exactly two excitations — not 40 particles.

Exact values for this step
SiteNoneMediumStrong
10 %0 %0 %
20 %0 %0 %
30 %0 %0 %
40 %0 %0 %
50 %0 %0 %
60 %0 %0 %
70 %0 %0 %
80 %0 %0 %
90 %0 %0 %
100 %0 %0 %
110 %0 %0 %
120 %0 %0 %
130 %0 %0 %
140 %0 %0 %
150 %0 %0 %
160 %0 %0 %
170 %0 %0 %
180 %0 %0 %
190 %0 %0 %
20100 %100 %100 %
21100 %100 %100 %
220 %0 %0 %
230 %0 %0 %
240 %0 %0 %
250 %0 %0 %
260 %0 %0 %
270 %0 %0 %
280 %0 %0 %
290 %0 %0 %
300 %0 %0 %
310 %0 %0 %
320 %0 %0 %
330 %0 %0 %
340 %0 %0 %
350 %0 %0 %
360 %0 %0 %
370 %0 %0 %
380 %0 %0 %
390 %0 %0 %
400 %0 %0 %
Now live · runs in your browser

Turn the dials yourself.

You choose the circuit settings. Your browser calculates the model probabilities in double precision, without a server or queue. The measured computation time appears below; it depends on your device and settings.

A particle starts in the middle and sets off as a wave in both directions at once. Twice the time, twice the width: 9.6 sites after 20 steps, 19.1 after 40 steps, 38.2 after 80 steps.

The interaction only acts between two particles.

Computed live · 161 sites · 80 stepsStep 80 / 80
Step 0more intense = more likely
Space-time picture of the position probability.
Site 1Start in the middleSite 161
Width of the cloud after 20 steps9.6 sites
Width of the cloud at the end38.2 sites
Still neighbours at the end–
Compute time in your browser–
Size does not matter

A million sites – and the middle stays the same.

Each gate can pass the wave on by only one site, so each step moves it by at most two. After 80 steps, everything farther than 160 sites from the middle is still exactly zero. Whether the chain has 401 or 1,000,001 sites therefore makes no difference in the middle. Check it yourself: your browser computes both chains completely, with one particle and your settings for tilt, disorder and looking.

What exactly does your browser compute here?

An open chain of 161 sites with one particle or 41 sites with two particles. Each step consists of two layers of the same two-site gate, as in the recordings above. Tilt and disorder then rotate the phase on every site. When you look after every gate, probabilities instead of amplitudes are carried forward. That is the same path without quantum effects.

The calculation is exact, with complex numbers in double precision – no random measurements, no approximation. With one particle there are 161 amplitudes, with two particles 820. That is why it is so fast. With many particles the effort grows rapidly; that needs computers like our i9.

The effects shown are well known and well studied in physics. We did not discover them. But here you can recompute them exactly yourself. This does not involve any quantum advantage.

Computed in advance on the i9 · 96 particles

Dam break: 96 particles that feel each other.

On the left all 96 sites of a chain are occupied, on the right all 96 are empty. Then the dam in the middle opens. Unlike above, many particles feel each other here at the same time. The number of their possible arrangements has 57 digits – no browser can go through them state by state. Our i9 computed it in advance with an established method; you replay the verified recording.

Loading the recording …

Newly computed on the i9 · 19 September 2026

Three ways a dam breaks.

Full on the left, empty on the right, then the dam opens – as above. Only the interaction changes. At strength 2 our model hits exactly the so-called Heisenberg point. Google Quantum AI studied this case in 2024 on a quantum processor with 46 qubits. Our i9 shows the same peculiarity.

Loading the curves …

What am I seeing, and why is it useful?

You are replaying actual results computed by our research prototype. Nine simulations cover three sizes and three interactions, each with 33 stored states. The animation is a player, not a live calculation or a measurement of physical time.

The bars show probabilities, not particle paths. A single measurement would occupy exactly two sites. Many measurements would reveal the displayed distribution. A stronger interaction need not produce a larger change at every step.

Models like this help study transport and interactions in quantum systems. This is an idealised, simplified model without noise. It is not a particular molecule or a medical simulation.

We computed these recordings on 14 September 2026 using the Intel Core i9-10900K CPU in double precision. We checked all 297 states against ffsim (established comparison software for these quantum calculations). The stored probabilities are rounded to six decimal places. Your selection stays in the browser; the page does not load our internal computational methods.

Another example · planned

Explore a molecule in the simulator.

We are planning a molecular lab for two hydrogen atoms. You will be able to vary their separation and use an energy curve to explore where the model describes a stable bond.

Today you can try the combination lock and our particle models. Before releasing the planned molecular lab, we will check its energy curve against an independent reference.

Explore the particle modelProject & progress
Our research standard: less work, equal accuracy

We investigate which quantum programs our simulator can run with less time or memory. That requires identical tasks, hardware and accuracy, plus repeated measurements. A larger qubit count alone is not evidence of better performance.

Our Hutter compression work motivates us, but text compression and quantum simulation are different tasks. These particle tests are an initial step towards more compact computation, not proof of a new general compression method.

The molecular lab is in planning. We will start with H₂ and name the basis functions used to describe the electrons. This small model explains a chemical bond; it cannot establish conclusions about cancer cells or drug effectiveness.

We publish measurements and test conditions. We do not release the source code of our internal optimisations. Claims of a breakthrough or a general speed advantage require independent comparisons.