Choosing a ComfyUI compute route when an RTX 5090 is not available
An RTX 5090 is one compute option, not a prerequisite for every ComfyUI project. Start with the required deliverable and compare the hardware and service routes that can produce it, keeping memory capacity, installation support and total operating effort separate.
Define the job before the device
Specify the model family, input type, output dimensions, number of deliverables and acceptable completion time. A still-image draft, a long video and a training task have different requirements. Do not use the requirements of the largest imaginable model as the baseline for a smaller actual project.
NVIDIA lists 32GB GDDR7 for the RTX 5090. That specification describes the device, not a guarantee that an arbitrary graph will fit. Model components, working tensors and optional processing stages all matter. System RAM and shared-memory figures should not be counted as additional physical VRAM.
Test the workstation you already have
Run a maintained minimal workflow in a separate, recoverable environment. Record whether the bottleneck is installation compatibility, model loading, sampling or decoding. A smaller trial output can answer a compatibility question, but it should not be reported as satisfying the final delivery specification.
Evaluate supported memory-saving options individually. If your current machine completes the required job reliably, a new purchase may not address the actual constraint. If it only completes after unacceptable reductions in output quality or size, keep that failure visible in the comparison.
Compare supported alternatives
For another local GPU, check the exact device, operating system and backend against the maintained ComfyUI system requirements. Broad labels such as NVIDIA, AMD or Intel are not sufficient compatibility evidence. The installer route may also differ between otherwise supported devices.
For remote execution, test your exact workflow rather than a provider's unrelated demonstration. Verify model availability, extension policy, transfer time, storage persistence and the process for retrieving both outputs and workflow files. Use non-sensitive trial inputs until data handling has been reviewed.
Make the comparison auditable
Keep a decision sheet with setup effort, accepted outputs, failures, measured waiting time and total charges for each candidate route. Obtain current quotations directly; this article deliberately does not publish prices that can become stale. Include recurring costs and operational work, not only the accelerator's headline price.
Choose the route that meets the present workload with an explicit contingency for growth. Retain the working baseline and its environment record so a later hardware change can be assessed against it. No alternative is declared universally faster or cheaper, and no unsourced benchmark is used to rank the options.