RTX 5090 in ComfyUI: 32GB VRAM versus 128GB system RAM

Check which memory number you are reading

NVIDIA's RTX 5090 specification lists 32GB of GDDR7. System RAM is installed on the motherboard and serves a different role. A seller's workstation description may list both capacities, so read the graphics-card line separately from the memory line before comparing configurations.

Record the exact installed board, driver and operating system. In a memory report, distinguish dedicated GPU memory, system memory and any shared-memory figure. Do not add the numbers together and describe the sum as VRAM. A workload using offloading can involve both memory pools without making them one physically interchangeable resource.

Locate the failing stage

First reproduce the problem using a maintained workflow and save its error log. Note whether failure happens while loading weights, encoding text, sampling, decoding or processing an upscale. These stages are not interchangeable. A sampler setting is unlikely to explain a failure before the sampler runs, and a VAE decode problem should not automatically be diagnosed as a missing checkpoint.

Use a small investigation record: workflow file, model filenames and hashes, input dimensions, batch size, precision choices, node versions, failing node and exact error text. Preserve the last successful configuration too. This gives you a concrete comparison rather than a general statement that a model is too large for the card.

Choose a memory experiment

The official troubleshooting guide documents memory-related options, including low-VRAM operation and preview configuration. Consult the options supported by your installed version. Offloading moves work or data between devices; it does not manufacture additional physical GPU memory.

• If several applications use the GPU, record the idle baseline after closing unrelated work and rerun the same graph.
• If sampling is the failing stage, try a smaller batch or lower input dimensions while leaving the model unchanged. Record the output-quality cost alongside whether it completes.
• If decoding is the failing stage, compare a documented tiled decode path only when the model and workflow support it.
• If weight loading fails, check the provider's supported precision and file variants. Do not rename a file or invent a quantisation format to make it appear compatible.

Change one factor at a time. A graph that runs after five simultaneous edits tells you that the combined configuration works, but not which change resolved the problem.

Measure capacity without inventing a benchmark

Keep capacity, peak usage and elapsed time in separate columns. A successful run establishes that one recorded configuration completed; it does not establish a universal maximum resolution or a speed advantage over another GPU. Repeat the same configuration and retain the outputs before publishing comparisons.

PyTorch's peak allocated-memory API measures tensor allocations tracked by that allocator. It is not a substitute for total device usage. Name the measurement tool and sampling interval, and do not present allocated and device-wide memory as if they measured exactly the same thing.

The 42 UK benchmark register defines the information required for a reproducible result. This article supplies no timing table because no qualifying RTX 5090 run is attached to it.

When 128GB of system RAM is relevant

A large system-memory configuration may be part of an offloading workflow, but capacity alone cannot establish its usefulness for your graph. Observe host-memory use, transfer time and completion time in the actual workload. Decide whether the slower successful run meets your delivery requirement rather than buying against an unverified headline.

For a purchase or migration decision, request a demonstration with the same model, output dimensions and workflow. Keep the test record with the quotation so the claimed capability has a specific meaning. Start from the installation matrix and maintained workflow references before treating a hardware upgrade as the only remedy.