Verdict
Two machines with the same 24 GB of video memory, built five years apart: Sam's desktop with an RTX 3090 from 2020, and the student's MSI laptop with an RTX 5090 Laptop GPU from 2025.
🟡 For today's SDXL work: about even. Her laptop pulls ahead on FP8 / FP4 models; Sam's desktop is better at running flat out for hours.
Same ceiling: both have 24 GB of VRAM, so every model that fits on one fits on the other (SDXL, Flux, LoRA training).
Same "engine size": both have 10,496 CUDA cores, and memory bandwidth differs by only 4% (936 vs 896 GB/s).
Her advantage is newer, not bigger: 5th-generation Tensor cores with hardware FP8 and FP4, and a cache 10× larger. It reaches 3090-level speed on 175 W instead of 390 W.
Sam's advantage is power and cooling: a desktop holds full speed indefinitely. A laptop running at 175 W can slow down during long batches as it heats up.
Not benchmarked: neither machine has been timed on the same job yet. The test in the last section settles it in 10 minutes.
The two machines
Sam's desktop
- GPU: NVIDIA GeForce RTX 3090 (desktop), 24 GB GDDR6X
- GPU power limit: 390 W (card allows up to 480 W)
- CPU: AMD Ryzen 7 5800X, 8 cores / 16 threads
- RAM: 64 GB DDR4-3200 (2 × 32 GB)
- Storage: 5 drives, about 10 TB raw (Samsung 990 Pro 1 TB, 2 TB SSD, 3 HDDs)
- Image stack: Forge, PyTorch 2.3.1 + CUDA 12.1
Student's laptop
- GPU: NVIDIA GeForce RTX 5090 Laptop GPU, 24 GB GDDR7
- GPU power limit: up to 175 W MSI spec
- CPU: Intel Core Ultra 9 285HX (24 cores) core count = Intel spec
- RAM: 96 GB DDR5-6400
- Storage: 5.59 TB (3.11 TB used)
- Also has: Intel integrated graphics (128 MB), used for the display only, not for AI work
GPU head-to-head
The figures that decide image-generation speed. Click any cell for an explanation.
| Spec | Sam: RTX 3090 desktop | Student: RTX 5090 Laptop | Edge |
|---|---|---|---|
| Architecture / year | Ampere (GA102), 2020 | Blackwell (GB203), 2025 | Student |
| Manufacturing process | Samsung 8 nm | TSMC 4N (≈ 5 nm class) | Student |
| Video memory (VRAM) | 24 GB GDDR6X measured | 24 GB GDDR7 screenshot | Tie |
| CUDA cores | 10,496 | 10,496 | Tie |
| Memory bus / bandwidth | 384-bit · ≈ 936 GB/s | 256-bit · ≈ 896 GB/s | Tie (Sam +4%) |
| L2 cache | 6 MB | 64 MB | Student |
| Tensor cores (AI math units) | 3rd generation | 5th generation | Student |
| FP8 in hardware | No | Yes | Student |
| FP4 (NVFP4) in hardware | No | Yes | Student |
| GPU power | 390 W limit measured | up to 175 W spec | Sam (sustained) |
| PCIe link | Gen 4 ×16 measured | Gen 5 ×16 spec | Irrelevant for this work |
| Upgradeable later | Yes, swap the card | No, GPU is soldered | Sam |
Unmarked rows are published NVIDIA specifications for each GPU model.
Per workload
Estimated relative speed, with Sam's RTX 3090 = 100. estimates from architecture and published benchmark patterns. Nothing here was timed on either machine.
Diana / Bella pipelines in Forge: ControlNet, ADetailer, LoRAs
Diana Fast: short jobs, little time to heat up
The 3090 has to convert FP8 back to 16-bit to compute
The newest compressed checkpoints
Long full load, where cooling decides
Sustained output, noise, heat
Parts are offloaded to system RAM
Rest of the system
| Part | Sam | Student | Matters for image generation? |
|---|---|---|---|
| CPU | Ryzen 7 5800X · 8C/16T measured | Core Ultra 9 285HX · 24C screenshot | Barely: the GPU does the work |
| System RAM | 64 GB DDR4-3200 measured | 96 GB DDR5-6400 screenshot | Only when models overflow VRAM, or for large video models |
| Storage | ≈ 10 TB across 5 drives measured | 5.59 TB (2.48 TB free) screenshot | Model library size and load times |
| Windows | Windows 10 Pro (build 19045) | Windows 11 Pro 25H2 (build 26200) | No |
Practical differences
⚠ Software on her laptop
- Install a current Forge or ComfyUI build made for CUDA 12.8 (labelled "cu128").
- Copy only the models, LoRAs and workflow presets across from Sam's setup, not the program folder.
- Update the NVIDIA driver to the latest Studio or Game Ready version.
- Once it runs, use FP8 versions of Flux and other new models. That's where her card is ahead.
How to settle it: a real test
Run the same job on both machines and compare seconds per image.
- Same checkpoint (e.g. the SDXL model Sam uses for Diana HD).
- Same prompt, seed 12345, 1024 × 1024, 30 steps, DPM++ 2M Karras, CFG 5, no ControlNet, no ADetailer.
- Generate once as a warm-up and discard it, then generate 10 images in a row.
- Note the "Time taken" Forge prints for the 10, and divide by 10.
- Repeat with batch 50 to see whether the laptop slows down once it is hot.
To capture her real GPU power and clocks, she runs this in a terminal while generating:
nvidia-smi --query-gpu=name,memory.total,power.draw,power.limit,clocks.gr,temperature.gpu --format=csv -l 2