Hello, today I will be presenting benchmarks for quantum computer simulations using the latest NVIDIA enterprise GPUs.
First, let's compare the performance between a CPU and the T4 GPU using the quantum volume benchmark.
こんにちは、本日は最新のNVIDIAの業務用GPUを使った量子コンピュータのシミュレーションのベンチマークを掲載します。
まずはCPUとT4 GPUを比較してみます。量子ボリュームのベンチマークを使います。
Quantum Volume
depth=30
shots=10
CPU / Qiskit
T4 GPU / Qiskit + cuQuantum, cuStateVec
H100 GPU / Qiskit + cuQuantum, cuStateVec
RTX6000ada GPU / Qiskit + cuQuantum, cuStateVec
We compared the results for 15-24 qubits using both a CPU and the T4 GPU. Even at this level, there is a noticeable difference.
15-24量子ビットの結果をCPUとT4 GPUを利用して比較しました。だいぶこの程度でも差が出ます。
CPU (sec.)
[0.05527604, 0.127417613, 0.269225495, 0.695744472, 2.309896906, 2.630688224, 5.06164711, 14.880064933, 27.564670069, 55.680368074]
GPU (sec.)
[0.027790243, 0.018253229, 0.029684344, 0.053728839, 0.100438104, 0.196138116, 0.172535676, 0.326441277, 0.643293149, 1.422738469]
Next, we compared the same calculations between the T4 GPU , 6000ada and H100. The T4 was able to compute up to 29 qubits, while the H100 and 6000ada could handle up to 31 qubits.
次に同じ計算をT4 GPUと6000ada, H100で比較してみます。T4は29量子ビットまで、6000adaとH100は31量子ビットまで計算できました。
T4 GPU (sec, 15 to 29qubits)
[0.02211307,
0.019795667,
0.028517052,
0.053816821,
0.100187135,
0.165203491,
0.18583667,
0.353295851,
0.702256174,
1.574080063,
3.140908516,
6.729538202,
13.634168086,
29.230031644,
59.903638174]
H100 GPU (sec, 15 to 31qubits)
[0.004408282,
0.002088285,
0.002336543,
0.003523071,
0.005282595,
0.008617269,
0.020409294,
0.038321433,
0.085984668,
0.172407263,
0.374064169,
0.751794014,
1.646756666,
3.367948447,
7.20650965,
14.179322945]
RTX6000ada GPU (sec, 15 to 31qubits)
[0.003902309,
0.004174494,
0.013700943,
0.006497565,
0.009483768,
0.017166366,
0.02926088,
0.060440877,
0.118379787,
0.238534027,
0.475478445,
1.02003665,
2.055512626,
4.442880517,
9.007608282,
19.095433893,
37.670469698]
Even a single H100 is sufficiently fast. I'll try it with multiple units when I have time. That's all for now.
単体のH100でも十分早いですね。時間がある時に複数台でやってみますね。以上です。