40 Qubit Simulation

[silicofeller]

We figured out that we need 64 H200 GPUs for a 40 Qubit Simulation. So, we have three questions for the Nvidia Team:

1. How many GH200 are required for a 40 Qubit simulation?
2. How many GB200 are required for a 40 Qubit Simulation?
3. How many B200 are required for a 40 Qubit Simulation?

Thanks,
Team Silicofeller

[yangcal]

Hello, are you referring to quantum circuit simulation with state vector method?

[silicofeller]

Can only cusvaer do this? Can we run cpu + gpu inside cuquantum appliance only? We used qiskit aer gpu for 37 qubit simulation on 8xH200 GPUs without cuquantum appliance. How do we get it to run a 38 qubit simulation ( which you told is possible if we use cpu memory)?

[tlubowe]

cusvaer, and CUDA-Q with the nvidia-mgpu backend can do this. Until IBM integrates the appropriate cuQuantum APIs only the stacks which have adopted this feature can support this.

[silicofeller]

Hi @tlubowe, hope you are doing well. Just to reiterate 16 Nvidia H200 GPUs can run a 40 qubit simulation using cusvaer by combining cpu and gpu memory. Right?

[tlubowe]

It depends entirely on how much memory each CPU has. If each Intel chip has 2.5TB of memory you should be fine. Let us know how it goes.

[silicofeller]

Hi Tom,

We have got the system with us. Currently it only runs one node. We have opened the cuquantum appliance but the simulations are not distributed to second node.

What should we do?

Can you create a quick youtube video to help us with running a multi node multi GPU simulation from start to finish. It will not only help us but the rest of the cuquantum community as well. It does not have to be too fancy at this point. Just a simple one will do or else even a zoom call will be very helpful.

We have tried almost everything even enroot. Need help on this.

Thanks!

[silicofeller]

We have successfully tested that 37 qubits state vector can be simulated using 8H200 GPUs. Also, we have not used cuquantum appliance for that but qiskit-aer-gpu for it. The problem is that we are not able to create a docker container to run cuquantum as the whole gpu pod that we are accessing online itself is a docker container. So cusvaer is not able, maybe if we use cusvaer we may be able to simulate 38 qubits using 8xH200 GPUs. We don't have access to GB200 now nor GH200 so we cannot figure out the exact number. If you could tell me how many GB200 or H200 are required to simulate 40 qubits successfully, you would have answered my question.

[TysonRayJones]

Hi there - I can help you estimate for deployment with cuStateVec.

EDIT: ah I only just noticed you were seeking to hybridise RAM and VRAM - my comments hare might be irrelevant!

A 40-qubit simulation maintains 2^40 complex amplitudes, each of which could be instantiated with single or double precision scalars (respectively requiring B = 4*2 or B = 8*2 bytes per amplitude, typically). So your statevector alone must be B 2^40 bytes which is 8 TiB or 16 TiB respectively. We will actually need some additional memory to act as a communication buffer when exchanging these amplitudes between GPUs, but we can bound it by x2, and I'm going to ignore it here completely.

This 8-16 TiB statevector must be partitioned between multiple GPUs, each of which can only instantiate a power-of-2 number of amplitudes. Each H200 GPU has 141 GB ~ 131 GiB of VRAM, and so can dedicate at most 128 GiB to instantiating the state - assuming nothing else of size ~3 GiB was needed in VRAM. As such, working in single or double precision respectively, you would need at least 64 or 128 H100 linked GPUs. Generally you can simply double costs for each additional qubit.