AMD and Qibo have showcased a new breakthrough in quantum simulation on classical hardware: an exact vector state simulation of 35 qubits run on a single AMD Instinct MI355X GPU. This achievement does not mean AMD has built a 35-qubit quantum computer, nor that practical quantum advantage has been reached, but it marks an important milestone for researchers needing to test algorithms, validate circuits, and study quantum systems before executing them on real hardware.
Exact quantum simulation is one of the most memory-intensive tasks. Each additional qubit doubles the size of the state vector, so jumping from 34 to 35 qubits is significant: it requires twice the memory and bandwidth to store and operate on all possible system states. Therefore, this news is more about the capacity of HBM memory, sustained performance, and the maturity of the ROCm stack than mere raw computational power.
Why 35 Qubits on a Single GPU Matters
In a state vector simulation, the classical system stores all the complex amplitudes describing the quantum circuit. For few qubits, this is manageable. Beyond 30, the problem quickly grows. With 35 qubits, we’re talking about 2^35 amplitudes—a scale that demands accelerators with substantial memory and very fast access.
AMD had previously demonstrated, alongside BlueQubit, simulations of up to 34 qubits on an Instinct MI300X. That GPU features 192 GB of HBM3 memory and 5.3 TB/s bandwidth. The new Instinct MI355X increases this to 288 GB of HBM3E and up to 8 TB/s, allowing for pushing the limits of exact simulation on a single card even further.
This “bit more” is meaningful because it eliminates, for certain tasks, the need to split quantum states across multiple GPUs. Multi-GPU simulations are necessary to scale further, but they add complexity: partitioning the state, device communication, synchronization, and potential efficiency loss. If a single GPU can handle a larger circuit, workflows become simpler and barriers for researchers and teams without access to large supercomputers are reduced.
This advancement was achieved with Qibo, an open-source quantum computing framework covering different layers: circuit simulation, benchmarking, validation, and hardware control. Its backend, Qibojit, employs JIT compilation and optimized kernels, with support for AMD GPUs via CuPy on ROCm. The goal is to enable researchers to work in a flexible environment and seamlessly transfer circuits between simulators and physical systems with fewer modifications.
MI355X, Qibojit, and the Role of HBM Memory
The comparison published by AMD focuses on Quantum Fourier Transform circuits, a common workload for gauging how a quantum simulator scales. The company states that the MI355X reduces total simulation time compared to the MI300X across the tested range, with greater improvements at higher qubit counts and larger memory pressure.
This highlights why data center GPUs are increasingly used beyond generative AI. The same combination of HBM memory, bandwidth, and parallelism that accelerates language models also benefits scientific simulation, molecular dynamics, numerical analysis, and here, quantum circuits.
AMD emphasizes that at many-qubit regimes, the bottleneck shifts from raw floating-point performance to memory access. It also notes that the jump from single to double precision remains contained in certain tests, though memory pressure escalates significantly when working with heavier representations.
It’s important to contextualize this. Simulating 35 qubits exactly on a GPU is useful, but it does not turn a GPU into a quantum computer. A classical simulator reproduces the mathematical behavior of a quantum circuit at an exponentially growing computational cost. A physical quantum computer, if sufficiently mature, should execute some problems differently. Until that hardware becomes more practical, simulators remain essential tools for algorithm design, circuit debugging, and preparing workloads for eventual real-system testing.
From Simulation to Quantum Hardware Control
The collaboration between AMD and Qibo extends beyond running simulations on GPUs. AMD’s blog highlights the roles of Qibolab and Qibosoq, key components of the Qibo ecosystem aimed at quantum hardware control. Qibolab converts circuits into pulses and control instructions, while Qibosoq enables communication with platforms based on QICK, an open-source system developed by Fermilab supported by AMD Zynq UltraScale+ RFSoC FPGAs.
This bridge between simulation and physical control is crucial. Many research teams want to design circuits, test them in simulators, and then transfer them to real hardware without rewriting all the logic. If the same framework facilitates movement between these phases, experimental development can be faster and less error-prone.
AMD aims to position itself with an open ecosystem strategy: Instinct GPUs for massive simulation, ROCm as a software platform, and RFSoC FPGAs for quantum system control and readout. This isn’t an attempt to directly produce a universal quantum computer but to provide part of the classical infrastructure supporting quantum computing.
Market-wise, while NVIDIA dominates much of the conversation around acceleration, AI, and scientific computing, AMD is gaining ground with abundant HBM memory, competitive prices, ROCm, and collaborations on open-source projects. In quantum simulation—where memory capacity is critical—a GPU like the MI355X offers a compelling option.
This access is further expanded via AMD Developer Cloud on DigitalOcean, offering instances with MI300X and MI350X accelerators. The aim is for researchers and developers to test Qibo on AMD platforms without owning hardware upfront. Although the setup still demands technical expertise, dependencies, and working with ROCm, it paves the way toward more accessible quantum software experimentation for laboratories and teams.
Real quantum computing still faces monumental challenges: error correction, noise, stability, physical scaling, and control. Meanwhile, classical simulation remains an essential tool. AMD and Qibo’s progress demonstrates that advancement depends not just on new qubits, but also on better ways to study and prepare algorithms via classical computing.
Frequently Asked Questions
What have AMD and Qibo achieved?
They executed an exact vector state simulation of 35 qubits on a single AMD Instinct MI355X GPU using Qibo and its backend Qibojit.
Does this mean AMD has a 35-qubit quantum computer?
No. It’s a classical simulation of a quantum system, not a physical quantum computer. It’s useful for testing algorithms and circuits, but the calculation runs on a conventional GPU.
Why is HBM memory so crucial?
Because the size of the state vector doubles with each added qubit. Beyond 30 qubits, memory capacity and bandwidth become as important as raw computational performance.
What role does Qibo play?
Qibo is an open-source quantum computing framework. It enables circuit simulation, algorithm testing, and integration with quantum hardware control tools like Qibolab and Qibosoq.
via: amd