cd /news/machine-learning/show-hn-tpu-accelerated-quantum-circ… · home topics machine-learning article
[ARTICLE · art-64696] src=github.com ↗ pub= topic=machine-learning verified=true sentiment=↑ positive

Show HN: TPU-accelerated quantum circuit simulation in Jax

A developer released a high-performance quantum circuit simulator built entirely in JAX, capable of simulating 36-qubit circuits with a 549 GB state-vector footprint on Google Cloud TPU v6e-64 clusters. The open-source project achieves ~0.01ms per gate and supports reverse-mode auto-differentiation, making it suitable for variational quantum algorithms. It is supported by Google's TPU Research Cloud program.

read2 min views1 publishedJul 18, 2026
Show HN: TPU-accelerated quantum circuit simulation in Jax
Image: source

Simulating 36-qubit quantum circuits at 549 GB scale. ~0.01ms per gate. 100% pure JAX.Accelerated on NVIDIA GPUs and Google Cloud TPU v6e-64 / v5e clusters. Supported by the Google TPU Research Cloud (TRC) program.

A high-performance, research-grade quantum state-vector simulator built purely in JAX. Run differentiable, noise-resilient, and large-scale quantum circuits accelerated on local NVIDIA GPUs and multi-worker Google Cloud TPU clusters.

100% Pure JAX: Zero dependencies on heavy frameworks (Qiskit, Cirq, Pennylane). Compiled natively into a single monolithic XLA kernel for bare-metal execution speeds. - Multi-Device Sharding: Scale up to 36 qubits (549 GB state-vector footprint) using distributed JAXPositionalSharding

across a 64-chip Cloud TPU v6e mesh. - Reverse-Mode Auto-Differentiation: Compute exact gradients in a single backward pass viajax.grad

for fast training of variational algorithms (VQE, QAOA, QNNs). - Hardware-Level Optimizations: Structured loop primitives (jax.lax.fori_loop

) prevent XLA graph bloat, whilejax.checkpoint

(gradient rematerialization) keeps memory complexity at$\mathcal{O}(1)$ . - Stochastic Noise Support: Built-in Monte Carlo trajectory simulations for open systems and depolarizing NISQ gate noise.

.
├── gpu/                     # GPU Modular Simulator & Research scripts
│   ├── jax_qsim/            # Core contraction engine (tensordot + transpose)
│   └── quantum_research/    # VQE, QAOA, GHZ state prep, noise trajectories
├── tpu/                     # TPU Scaling Suite (experiments and runners)
├── shors/                   # TPU-sharded Shor's Algorithm (33 qubits)
├── grover_simulation/       # Grover's Search (up to 36 qubits on 64 TPU chips)
├── tests/                   # Pytest verification suite
└── requirements.txt         # Core dependencies

Ensure you have CUDA 12 installed, then set up the environment:

python3 -m venv venv && source venv/bin/activate
pip install --upgrade "jax[cuda12]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
pip install matplotlib pytest numpy

Verify the JAX device setup:

python3 -c "import jax; print('Backend:', jax.default_backend()); print('Devices:', jax.devices())"

Run the local GPU benchmarks:

python benchmarks/benchmark_27q.py

In your TPU VM cluster SSH session:

python3 -m venv tpu_env && source tpu_env/bin/activate
pip install "jax[tpu]" -f https://storage.googleapis.com/jax-releases/libtpu_releases.html
pip install matplotlib numpy

Run the scaling suite:

python tpu/tpu_quantum_scale.py
Environment Hardware Max Qubits State-Vector Footprint Gate Speed (10-q)
Local GPU
NVIDIA RTX 2050 (4 GB VRAM) 29 ~4.29 GB ~0.01 ms
TPU Mesh (v5e-16)
16x TPU v5e (256 GB aggregate HBM2e) 33 64.00 GB ~0.01 ms
TPU Mesh (v6e-64)
64x TPU v6e (2.0 TB aggregate HBM3) 36 549.76 GB ~0.01 ms

We are extremely grateful to the TPU Research Cloud (TRC) program by Google for providing access to Cloud TPU v6e and v5e VM clusters that enabled this scale of research.

Licensed under the Apache License 2.0.

── more in #machine-learning 4 stories · sorted by recency
── more on @google 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/show-hn-tpu-accelera…] indexed:0 read:2min 2026-07-18 ·