Why Parallelize?
Why Parallelize?
Quantum workloads are many circuits
Most practical quantum algorithms run hundreds or thousands of circuits:
| Workload | Where the circuits come from |
|---|---|
| VQE and other variational algorithms | Each optimiser step evaluates the cost on one or more parameter sets, and each evaluation needs one circuit per measurement group. Gradient methods add shifted copies. |
| QAOA on large problems | The problem is partitioned into sub-problems, each running its own QAOA instance. |
| Parameter and geometry sweeps | One program per point, for example per bond length in a potential-energy curve. |
| Time evolution | One circuit per time point, plus repeated randomised samples for QDrift. |
| Error mitigation | Zero-noise extrapolation and Pauli twirling multiply each circuit into several variants. |
Run one at a time, these circuits leave most of the available hardware idle and expose the user to every queue delay.
Where the stack parallelises
In Divi (client side)
- Ensembles. A partitioned problem, a sweep, or a trajectory becomes a program ensemble: many independent programs that advance together, round by round.
- Batch merging. Divi’s batch coordinator collects the circuits that all programs in an ensemble need for their current step and submits them as one large job instead of many small ones.
- Server-side work. On the cloud, expectation values are computed by the simulator and parametric circuits are bound on the server. This keeps the job compact and avoids measurement-basis blow-up.
In Composer (cloud side)
- Simulator fleet. A job’s circuits are split into cost-balanced batches that are processed concurrently by an autoscaled pool of Maestro workers.
- QPU systems. A QPU system can contain several devices. Circuits are spread across every device in the system that is online.
- Circuit packing. Small circuits are placed side by side on one QPU so that a single device run serves several circuits.
In Maestro
- Batches, shots, and noise realisations run in parallel on multicore CPUs. GPU backends accelerate wide statevector, tensor-network, and density-matrix simulations. Very large statevectors can be distributed over several GPUs or nodes.
flowchart LR
A["Divi ensemble<br/>(many programs)"] --> B["Batch coordinator<br/>one merged job"]
B --> C["Composer<br/>cost-balanced batches / packed QPU runs"]
C --> D["Maestro worker"]
C --> E["Maestro worker"]
C --> F["Maestro worker"]
C --> G["QPU device(s)"]
D --> H["Results"]
E --> H
F --> H
G --> H
H --> I["Divi: next optimisation round"]
I --> A
A single job targets one simulator cluster or one QPU system. Mixing hardware types is done on the client side: Divi can send different programs to different backends.