What AI and quantum simulation can actually do on hardware you can own.
GoldenHammer Innovations runs independent research in GPU-accelerated quantum simulation, local agent systems, and edge inference. The common thread is measurement on real hardware: we build the system, run it, and publish what it did, including where it fell short.
Quantum statevector simulation on edge GPUs
The question behind this work is simple. On NVIDIA's Jetson Orin, an edge module where the GPU and CPU share one pool of memory, when does GPU acceleration actually beat the CPU for quantum circuit simulation?
Answering it meant first getting the GPU path to work at all. NVIDIA's prebuilt cuQuantum Python wheels ship without device code for the Orin's sm_87 architecture, and on that hardware the cuStateVec kernels silently skip their work without raising an error. We confirmed the cause three independent ways, through a direct cuStateVec API test, an inspection of the compiled binaries showing sm_87 absent, and a blocking-launch run that placed the failure in the library rather than the CUDA runtime. The fix was to build qiskit-aer 0.17.2 from source for compute capability 8.7.
With a working GPU path, the benchmarks map the crossover.
| Workload | Result | Why |
|---|---|---|
| Random circuits, 20 to 26 qubits | GPU, up to 1.55× | Faster at every size, peaking at 23 qubits and easing to about 1.3× at 25 and 26 as shared memory bandwidth saturates. |
| 16-qubit variational classifier, one shot at a time | CPU, 2.19× | At this size, kernel launch overhead outweighs the work the GPU saves. |
Where it stands. The manuscript is complete and a preprint is forthcoming. The next phase takes the simulation past the memory of a single device onto multi-node GH200-class systems, which is where we are looking for academic and compute partners.
ATHENA, a local agent platform
ATHENA is an agent system that runs on the lab's own hardware, built around a Jetson Orin Nano Super, rather than on a cloud model. It routes in two stages: a fine-tuned 3-billion-parameter model makes the first decision, choosing which tool to call, and a 14-billion-parameter model writes the answer from what the tools return.
The tool-calling model was trained on a 4,663-example multi-turn corpus produced by the system's own distillation pipeline, built so that no test material leaks into training. Retrieval runs over more than 37,000 indexed passages covering tax law, security, physics, and networking.
In the full 367-case evaluation, run end to end through the agent twice, it produced a complete answer on 97 to 98 percent of cases and fired the expected tool on roughly 87 to 88 percent. An audit of the misses traced most of them to the test set itself, such as questions that never named the machine they were about, so the honest estimate of true model error is lower than the raw miss rate.
Where it stands. ATHENA is at version 1.2. The next step is relabeling the evaluation set and rerunning it, so the published numbers measure the model rather than the test.
Edge inference and the lab
The standing question here is practical: which workloads can leave the data center for hardware a small company can own, and what do they cost and deliver when they do? Answering it takes hardware we can instrument end to end, so the lab was designed, built, and validated in-house.
The training node is a Threadripper 2950X workstation with three RTX 3090-class GPUs and 72 GB of combined VRAM, fed by two power supplies on separate circuits. It completed a three-GPU full-load burn validation at 1,007 watts peak with zero errors. The edge node is the Jetson Orin Nano Super that runs ATHENA and the quantum benchmarks, which is where the latency-sensitive inference work happens.
Physics
Two longer-horizon projects sit alongside the applied work. The first tests primordial black holes as dark matter and as the seeds of supermassive black holes, using nested sampling over a model of the early universe's curvature power spectrum. The second scopes distributed quantum sensing in deep space: networks of quantum sensors and synchronized clocks for positioning, navigation, and timing in cislunar space, where GPS is not available.
Where it stands. The black hole analysis has a draft manuscript and is still being tightened. The sensing work is in scoping.
Working together
The multi-node phase of the simulation work needs partners: academic collaborators and access to GH200-class compute. If you work in quantum simulation, high-performance computing, or edge AI and any of this overlaps with yours, write to adam@g-h-i.com.
Adam Mitchell, ORCID 0009-0004-4028-4471.