BenedictThis is a reading guide to the version 0.1 preprint. The linked PDF contains the complete methods, measurements, limitations, and references.
A faster accelerator is not always a faster workload.
A local computer can reach far beyond the hardware attached to it. But making a remote GPU useful involves acquiring capacity, transferring data, starting a runtime, collecting artifacts, and continuing the work locally. Those costs can outweigh the computational advantage.
Naught Fusion Alpha explores that boundary: when should an external resource participate in a workload, and how can the runtime preserve the same computation across machines?
Explicit stages. Verified state.
The prototype represents workloads as dependency graphs. It first filters resources by capability, then makes deterministic placement decisions using measured or profiled computation cost, object locality, transfers, startup, and coordination.
Intermediate data is stored as immutable, content-addressed objects. Run identifiers, byte lengths, formats, and SHA-256 checks connect the output of one stage to the input of the next. Compact control messages, bulk data, and executable payloads are staged separately.
This is explicit workload staging, rather than automatic partitioning of arbitrary source code.
What the alpha shows
- Small workloads can lose badly. A small-kernel microbenchmark took 42.0 milliseconds locally and 118.3 seconds through a managed NVIDIA L4 path.
- Heavier work changes the comparison. An isolated external kernel-family profiling run took 134.0 seconds, against roughly 145–147 seconds locally. This proxy generated its input remotely and excluded Fusion’s real input transport.
- A complete heterogeneous path was verified. Local Stage A produced an immutable object; the exact object was staged and verified remotely; an NVIDIA GeForce RTX 5090 executed CUDA Stage B; and local Stage C consumed the returned artifact and passed correctness checks.
Feasibility, with boundaries.
The first successful run’s original timer included a post-run local reference computation, and some staging was not separately timed. The paper explicitly separates that result from a performance claim.
The evidence comes from an alpha prototype, one verified complete heterogeneous success, and a small number of real GPU measurements. Resource readiness, hardware substitution, transfer estimates, and failed or cancelled attempts all affect what can be concluded.
The central finding is architectural: the value of external acceleration depends on the complete path required to acquire, reach, and use it.
Continue with the full paper
Naught Fusion: Workload-Level Fusion of Local and Remote Heterogeneous Compute. Benedict Blessing Gbogr. Version 0.1 preprint, 8 Sept 2026.
Open the 8-page PDF