JUPITER supercomputer hits 50 qubit simulation record using exascale CPU GPU hybrid - honest opinions

[Coder53]

Jülich and NVIDIA pulled off the first full universal simulation of a 50 qubit quantum computer using JUPITER, Europe's first exascale machine. The previous record was 48 qubits set on Japan's K computer back in 2019, so this is a meaningful jump given how the memory requirements scale exponentially with each added qubit.

The point of these classical simulations is to validate algorithms like VQE and QAOA before you run them on real hardware that still has unacceptable error rates. It also gives you a benchmark for when actual quantum hardware actually crosses the classical simulability threshold for arbitrary circuits.

What I find more interesting is the heterogeneous CPU GPU architecture exploitation. The paper on arXiv goes into the load balancing between the partitions and it is genuinely clever work. The arms race between classical simulators and real quantum hardware is producing better tools on both sides

JUPITER supercomputer breaks world record with 50-qubit quantum simulationwww.sciencedaily.com

[Upsilon]

50 qubits classically simulated is impressive but lets be real this is petabytes of state vector

[Taker04]

Per the paper around 36 petabytes for the full state, yeah this is not your laptop running Qiskit

[Shane96]

The whole point of quantum advantage is doing what classical cannot, so simulators getting better is the wrong framing here

[Drifter]

Disagree, better classical simulators raise the bar for what counts as genuine quantum advantage which is healthy for the field

[CrimsonFury]

Anyone know if they were running full universal gate sets or restricted Clifford plus T

[GlassyCandle]

Universal per the abstract, full state vector simulation not stabiliser

[Inland Sienna]

JUPITER coming online has been a quiet revolution for European computational science, good to see it getting use

[CosmicRay67]

Why is it always Germany doing this kind of infrastructure work, the UK has been asleep at the wheel on HPC for a decade

[Steady Dylan]

Funding, mostly. EuroHPC actually puts money down where the UK just publishes strategy documents

[RayOfLight]

Curious what the energy cost of this simulation was, exascale runs are not cheap on the power bill