Why cuStateVec via Qsim with gpu_mode=(0,) slower than one with gpu_mode=1 for QFT circuit?

[rht]

I was exploring the gpu_mode option on a NVIDIA cuQuantum Appliance Docker instance.
I was testing the usual QFT circuit on QsimCirq, and I tried the 2 options

options = qsimcirq.QSimOptions(gpu_mode=(0,))
options = qsimcirq.QSimOptions(gpu_mode=1)
simulator = qsimcirq.QSimSimulator(options)

I found that gpu_mode=(0,) (multi-GPU backend but with only 1 GPU) to be consistently faster than the single-GPU backend.
My measurements for the elapsed time it takes to run the circuit, in seconds:

• QFT 28 qubits, gpu_mode=(0,): 1.10,1.11,1.09
• QFT 28 qubits, gpu_mode=1: 1.45,1.45,1.45
• QFT 30 qubits, gpu_mode=(0,): 3.67,3.80,3.59
• QFT 30 qubits, gpu_mode=1: 5.32,5.58,5.33

Why is this the case? Should I stay away from gpu_mode=1?

[leofang]

Hi @rht, thanks for sharing your observation. It's likely because there're extra optimizations that we offer with our custom mgpu backend, such as gate fusion. If you can use our cuQuantum Appliance container, it'd be better indeed if you use the mgpu backend even if you are on single GPU. But, for performance chasing games like this, it's always advised to benchmark with your workload first 🙂

[mtjrider]

As @leofang points out: in the appliance, gpu_mode=(0,) triggers the mulit-GPU backend running on a single GPU.
gpu_mode=1 will force the simulator to use only cuStateVec as a backend.

The multi-GPU backend contains additional optimizations for circuit-level simulation.

Here's the documentation link for reference.

[rht]

I see, my question is resolved. And I'm using cuStateVec for both options. I confirm that setting max_fused_diagonal_gate_size=0 for gpu_mode=(0,) gives about the same elapsed time as gpu_mode=1. An optional question is that why is the diagonal gate fusion not applied in the single GPU backend?