In a technical milestone announced at GTC Berlin 2026, NVIDIA and the Jülich Supercomputing Centre (JSC) at Forschungszentrum Jülich have executed a 104-qubit fault-tolerant quantum computing simulation on JUPITER, Europe’s first exascale supercomputer.
Detailed in an official release and accompanying NVIDIA technical report (“Advancing Fault-Tolerant Quantum Computing with GPU-Accelerated Simulation“), the simulation modeled the physical execution of magic state cultivation under realistic hardware noise—validating a fault-tolerant protocol that reduces physical qubit overhead by up to 90% compared to traditional cascade distillation factories.
| [ JUPITER Exascale Simulation & Hardware Infrastructure Profile ] | ||
|---|---|---|
| System / Software Pillar | Hardware & Computing Infrastructure | Simulation Scale & Performance Metrics |
| • Supercomputing Platform | • JUPITER (Jülich Supercomputing Centre) | • 8,192 NVIDIA GH200 Grace Hopper Superchips connected via Bull exascale interconnect. |
| • Development & Staging | • NVIDIA Accelerated Quantum Research Center (NVAQC) | • Staged on NVIDIA GB200 NVL72 Grace Blackwell racks to optimize 4-to-hundreds GPU scaling before exascale run. |
| • QEC Protocol Simulated | • Magic State Cultivation (104 Qubits) | • Distance-3 Color Code → Distance-7 Hybrid Color/Surface Code transition with active flag error detection. |
| • Noise & Physical Modeling | • Coherent Phase Noise + 3-Level Leakage | • 45-bit encoded state vector tracking ~35.2 trillion complex amplitudes across 512 TiB distributed GPU RAM. |
| • Software Libraries | • NVIDIA CUDA-Q™ + cuQuantum (cuStateVec) | • Packed qutrit-to-binary state mapping, cuBLASMp, and cuSOLVERMp distributed factorization. |
Exascale Simulation Mechanics & Qutrit Encoding
Simulating magic state cultivation requires tracking non-Clifford operations and non-Markovian noise that cannot be simplified by stabilizer tableau methods (such as the Gottesman-Knill theorem):
- Qutrit Leakage Encoding: To capture physical superconducting qubit errors where state population escapes the computational subspace (|0⟩, |1⟩) into a third energy level (|2⟩), the team packed groups of three-level qutrit systems into binary registers (e.g., 5 qutrits require 243 basis states, packed into 8 binary bits).
- Memory & Scale: The 45-bit encoded state vector tracked 35.2 trillion complex amplitudes, consuming 512 TiB of aggregate GPU memory (64 GiB per GH200 node) across 8,192 GPUs on JUPITER.
- Development Pathway at NVAQC: Development, verification, and code scaling were conducted at the newly established NVIDIA Accelerated Quantum Research Center (NVAQC) using GB200 NVL72 architectures, reducing simulation development timelines from weeks to days before launching on JUPITER.
The exascale simulation directly complements Google Quantum AI’s foundational cultivation theory, builds upon Google’s Willow hardware cultivation demonstration in Science, and expands NVIDIA’s CUDA-Q platform momentum across European supercomputing hubs.
Review the press announcement on the NVIDIA Newsroom here, examine the technical blog on the NVIDIA CUDA-Q Portal here, and inspect the NVAQC platform overview on NVIDIA Quantum Solutions here.
October 19, 2026