Auditing a claimed 48GB RTX 4090 benchmark
A screenshot showing a large memory figure is not enough to establish that a modified RTX 4090 works reliably. This guide explains the evidence a hardware-capacity claim needs, and how to separate physical device memory from system RAM, allocator accounting and an edited label.
Establish the manufacturer baseline
NVIDIA specifies 24GB GDDR6X for the standard RTX 4090. A claim about a different capacity therefore concerns a nonstandard device, not a hidden software setting on the retail card. Record the board model and the original specification before interpreting any reported upgrade. This article does not verify a particular modified board or provide a modification procedure.
Ask for a continuous record of the inspected machine and its environment rather than a detached benchmark chart. Device identifiers, driver versions and diagnostic output should refer to the same session. Redact serial numbers and account details before publishing the evidence.
Identify what the number measures
NVIDIA's management-interface documentation distinguishes device-memory reporting from other device statistics. PyTorch's peak allocated-memory function measures memory occupied by tensors through its allocator. Those are useful but different views; neither a Windows shared-memory total nor a process's system-RAM allocation establishes extra physical VRAM.
A useful result names the measurement tool, units and sampling interval. Distinguish bytes, decimal gigabytes and binary gibibytes. Preserve the complete output instead of cropping away the device name or the distinction between used, reserved and total memory.
Require a reproducible workload
The evidence package should identify the exact workflow, model hashes, input files, precision and output dimensions. Include startup logs, failures and the final saved output. A memory allocation demonstration does not establish successful image generation, and one generated image does not establish sustained reliability.
Compare the claimed device against an unmodified baseline using the same environment where possible. Document any unavoidable differences. Run repeated workloads and inspect outputs for corruption, not just elapsed time. Do not extrapolate an unverified result to training, video generation or a different model family.
Publish a defensible conclusion
Classify the result as manufacturer specification, third-party claim, independently reproduced observation or unresolved evidence. These labels stop an interesting demonstration becoming an unsupported buying recommendation. More memory can permit a larger workload without making an already-fitting workload faster.
The earlier version of this page presented upgrade benefits as established findings without a linked run record. That interpretation is withdrawn. Until a complete reproducible package exists, use the standard specification for capacity planning and treat any modification as an unverified hardware experiment, not a supported ComfyUI optimisation.