Lightning-AMDGPU Advances Quantum by AMD and PennyLane
Lightning-AMDGPU connects PennyLane quantum software to exascale HPC systems for fast, scalable quantum simulation on AMD Instinct GPUs.
AMD hardware powers High-Performance Computing (HPC), and PennyLane's quantum software ecosystem works well with it. AMD accelerators power the Top500 leaders of today and the upcoming exascale behemoths Frontier, El Capitan, Alice Recoque, Discovery, and Lux, which are changing the environment for accelerated scientific discovery. PennyLane facilitates quantum research on any hardware, featuring massive simulations on flagship supercomputers, AMD Developer Cloud access, and personal workstation prototypes.
Smooth Lightning-AMDGPU Access
PennyLane's Lightning-AMDGPU simplifies AMD GPU quantum simulation. This tool contains AMD GPU-specific Lightning-Kokkos simulator precompiled binaries. Developers can use AMD Instinct GPUs in the release, which is optimised for usability and performance, with a simple installation command on a Developer Cloud instance.
Lightning-AMDGPU uses the Kokkos portability framework-driven Lightning-Kokkos simulator backend. C++ code runs easily on CPUs and GPUs with this framework. The native programming model, HIP, of AMD processors automatically lowers this code for best performance. Precompiled Lightning-AMDGPU wheels are available for ROCm 7.0 and MI300 GPUs.
The AMD Developer Cloud makes high-end AMD GPUs like the MI300X more accessible without supercomputing capabilities. Developers can use this resource like a cloud provider by creating a GPU Droplet and selecting the MI300X plan with the AMD ROCm 7.1 Software image installed.
Frontier-validated MPI massive scalability
Accessibility is crucial, but scaling quantum workloads to the max advances research. With ORNL's help, PennyLane and the Lightning simulator were launched on Frontier, the first exascale supercomputer, proving its scalability.
Sometimes exascale computing is thought to need complex coding. However, Lightning on Frontier was easy to install and use, and thorough instructions helped customers maximise the system's high-bandwidth connectivity.
PennyLane Lightning is “HPC-friendly” and supports high-performance. Lightning-Kokkos supports MPI since 0.42. MPI provides the “secret sauce” for massive scaling by enabling programmers distribute the state vector of a circuit over several nodes and GPUs. Researchers can simulate more qubits than a GPU can handle and run large-scale simulations faster with this feature.
Lightning-Kokkos can replicate circuits on over 1000 AMD GPUs, where strong scaling graphs for running the Quantum Fourier Transform (QFT), a crucial subroutine in many quantum algorithms, show performance gains when increasing hardware resources.
Supercharger Hybrid Compilation
Beyond simulator scaling, complex hybrid quantum-classical operations must be optimised. PennyLane's Catalyst QJIT compiler meets this demand. Catalyst speeds up and enables AutoGraph and optimised dynamic quantum circuits by creating hybrid programs.
The Lightning-AMDGPU and Lightning-Kokkos backends should work well with Catalyst. This interaction lets customers use Lightning's raw throughput on an AMD ROCm device and Catalyst's compilation capabilities. This combination lets the entire workflow, not just the quantum circuit, be just-in-time assembled for optimal speed and offloaded to AMD GPUs.
Most importantly, Catalyst improves optimisation. As circuit gate depth increases, PennyLane takes longer to merge rotation gates and cancel successive adjoint gates to optimise a sample circuit. Catalyst can maintain a quantum optimisation control structure since its compilation time is constant regardless of gate depth.
Introduction to PennyLane and AMD
Developers can use PennyLane on AMD hardware currently in several simple ways:
Pip Install on AMD Developer Cloud: Installing PennyLane and pennylane-lightning-amdgpu with pip is the easiest way to run on MI300X GPUs.
Docker Images: Follow AMD's Container Toolkit quick start instructions before running the Docker command. We give pre-built pictures.
Building from Source: Use CMake, Ninja, and hipcc to build Lightning-AMDGPU or Lightning-Kokkos from source for optimal performance, bespoke configurations (like Frontier), or MPI scaling.
These methods let users run PennyLane programs like the QFT, Bell State Circuit, and Variational Quantum Circuit (VQC), which employs JAX for gradient computation. VQC and QFT benchmarks run faster on the GP (lightning.amdgpu) than the CPU (lightning.qubit).
