Nvidia’s homegrown combined CPU and GPU offering, called Grace Hopper, has been slow out of the gate, but the chipmaker is finding a new use for it in enabling hybrid quantum computing.
The Grace Hopper chip is being used alongside Quantum Machines’ quantum hardware to facilitate quantum-classical computing.
Nvidia is finding a good use of its homegrown Arm-based CPU, which will play the role of an intermediary between the quantum processors and a simulated quantum environment on the GPU.
Quantum computers are still under development, but researchers and infrastructure providers are recognizing that quantum computers will be an accelerator in traditional computing environments.
The idea is to break up quantum workloads so only the appropriate tasks go on to the quantum processors, and its operation is supplemented by GPUs. Nvidia has software tools capable of creating quantum-computing environments in GPUs.
Nvidia’s DGX Quantum system, which has the Grace CPU and Hopper GPU, is connected to the OPX+ quantum controller by Quantum Machines. The GPUs are running classical workloads to accelerate error correction, calibration, control and hybrid algorithms, a spokesman from Nvidia said.
For example, quantum error correction is offloaded to the GPUs, which is more of a classical computing workload.
Nvidia has created a tool called CUDA Quantum (formerly known as QODA) that mimics a quantum environment in GPUs. The OPX+ quantum controller is more of an interpreter that allows researchers to not worry about the type of qubit being used for processing.
Quantum companies are developing their own qubit, which have their own software toolkits and compilers. The different technologies and tools have created a diversified hardware and software environment.
IBM, Google and Rigetti are developing superconducting qubits, while Intel and PsiQuantum are focusing on quantum dots and silicon photonics, respectively.
Nvidia says its quantum control system is based on the CUDA Quantum open-source programming model. But Nvidia’s software is designed for acceleration of its homegrown GPUs. That means a user will have to either need to buy this hardware or look for one in the cloud, which can be pricey.
In the DGX Quantum system, any quantum workload goes to the Grace CPU, which in conjunction with the Quantum Machines controller facilitates the distribution of tasks between quantum systems and GPUs. The output goes back to the CPU, which then presents the results back to users.
This system fits into ongoing research of splitting quantum workloads into more classical computing environments or alternate circuits. It is considered more power-efficient as quantum computers do not need to run low-priority workloads.
Dell has plans to use a technology called quantum circuit-cutting in environments where its conventional servers link up to a quantum computer from IonQ, which uses trapped ion qubits. IBM has plans for a classical-quantum environment with dynamic circuits, which involves breaking workloads into smaller circuits.
The Grace CPU in DGX Quantum has a total of 144 cores based on Arm’s Neoverse V2 design.
The Grace Hopper multichip module connects the CPU to the Hopper GPU via the NVLink-C2C connector, which has a speed of 1 terabyte per second. It uses LPDDR5X memory.
Nvidia has introduced multiple systems and reference designs with the Grace Hopper Superchip, which include OVX for the metaverse and HGX for AI and high-performance computing. The systems include BlueField-3 board for network, storage and data processing.
The DGX Quantum system connects to OPX+ quantum controller via a PCIe slot. The Grace Hopper system supports the PCIe Gen5 interconnect.
Server makers Supermicro, Asus and Lenovo said they would release servers with Grace Hopper in the first half of this year. However, no third-party systems with Grace Hopper are yet available. Current systems from these vendors link up the H100 GPUs to x86 CPUs from Intel or AMD.
Header image: Nvidia’s DGX Quantum system