Practical research · explicit evidence
RTX 4090 VRAM: 24GB limits, 48GB claims and practical alternatives
The standard GeForce RTX 4090 has 24GB of GDDR6X. A report of a 48GB modified card is a claim about a particular altered board, not an upgrade setting or a guarantee that your ComfyUI workflow will run.
Check what is actually limiting the job
Record the card model and total memory reported by the driver. Keep the full error message and the stage where it happened. Model loading, sampling, decoding and upscaling can have different peaks. Available VRAM is also affected by the desktop and other applications.
A display reporting more memory is not evidence of stable operation across the whole address range. Equally, an out-of-memory error does not prove that replacing a card is the only solution. First make the target workload precise: model, graph, dimensions, batch size and required output quality.
Compare changes by what they give up
| Option | What changes | Evidence to collect |
|---|---|---|
| Reduce batch size | Fewer outputs are processed together. Total throughput may change. | Record successful outputs per elapsed minute, not only seconds for one batch. |
| Reduce dimensions or split optional stages | The workload or pipeline changes. A smaller result may not satisfy the original delivery requirement. | Preserve original and altered graphs and review the final output. |
| Use supported memory-management options | Some work or data may move between system memory and the GPU. | Measure end-to-end time, memory, RAM use and failure rate under fixed versions. |
| Use a GPU or service with adequate documented memory | Hardware, driver, cost or data location changes. | Run the actual graph at delivery settings before making a purchase or migration decision. |
| Consider an aftermarket board modification | Board-level hardware and support assumptions change. | Require evidence for the specific board and its workload; a capacity screenshot alone is insufficient. |
No row is automatically best. If the job must finish locally and overnight, a slower memory-saving mode might meet the requirement. If interactive turnaround is essential, the same trade-off might fail it. Evaluate the constraint you actually have.
What would make a 48GB claim useful?
- Identify the exact board and modification, with provenance rather than a stock product name.
- Show driver and operating-system versions and the reported usable capacity.
- Provide repeatable memory and workload tests, with duration, errors and output validation.
- Run the specific ComfyUI graph across repeated clean starts. Keep failed attempts in the record.
- Explain power, cooling and support arrangements for that altered hardware. Do not assume standard-board documentation covers it.
This is an evidence checklist, not a soldering guide. Hardware work requires appropriate equipment and expertise; the article does not establish that any particular modification is reliable.
A small decision exercise
Suppose a graph fails during the final upscale. Compare the original run with a separate upscale job, keeping the delivered dimensions and visual acceptance rule fixed. Record whether both stages finish, their combined time and their outputs. That observation is more useful than saying “24GB is not enough” without naming the graph. The example describes an experiment; it is not a result from our equipment.
Download the blank run-record CSV Full benchmark methodology
The CSV contains column headings only. It is a recording template, not a tested workflow or measured result.
Use the stage-by-stage RTX diagnostic guide before choosing a hardware route.
Primary specification
NVIDIA GeForce RTX 4090 specifications: standard memory configuration, 24GB GDDR6X. Checked 16 September 2026.