Choosing Your Machine

Pick the right CloudDock GPU in 60 seconds — based on VRAM, model size, and what you’re actually doing (not your ego).

Quick picker (60-second answer)

If you don’t want to read the whole page, use this:

  • A (RTX 3060 class) → best value for SD 1.5 image generation, light ControlNet, basic hires.
  • B (RTX 3090 class) → pick this for SDXL, heavier ControlNet stacks, higher-res hires, more “room to breathe”.
  • C (A100 40GB class) → for training (LoRA / heavier jobs), big batches, fewer OOM surprises.
  • D (A100 80GB class) → for large-scale training, big VRAM workloads, and “I really don’t want to babysit memory”.
Golden rule: If you’re just generating images, start with A. Upgrade to B only when you have a clear reason (SDXL / high-res / heavy ControlNet). Go C/D when you’re training or doing truly VRAM-heavy work.

What you’re actually paying for

Most “which GPU is better?” questions are actually “how much VRAM headroom do I need?” More VRAM means:

  • Higher base resolution without OOM
  • More ControlNet units at once
  • Higher hires scale at safer denoise
  • Bigger batches (or faster iteration per hour)
  • Training workflows that don’t explode

CloudDock groups at a glance

A/B/C/D comparison table for CloudDock machines
A/B/C/D: think of it as VRAM tiers + workflow tiers.

A/B/C/D comparison (practical view)

Tip: Numbers below are “what usually feels comfortable”, not hard limits. Your exact ceiling depends on checkpoint, ControlNet, batch size, and whether you’re doing hires/training.
  • A — RTX 3060 class (entry / best value)
    • Best for: SD 1.5 generation, anime art, basic ControlNet, light hires
    • Typical comfort: 512–900px base, hires ~1.2×–1.35×, ControlNet 1–2 units
    • Avoid: aggressive SDXL + heavy ControlNet + big hires all together
  • B — RTX 3090 class (SDXL / high-res comfort)
    • Best for: SDXL generation, higher res, heavier ControlNet stacks, faster iteration
    • Typical comfort: 768–1200px base, hires ~1.3×–1.6×, ControlNet 2–3 units
    • Good when: you’re tired of “almost works” on A
  • C — A100 40GB class (training + stability)
    • Best for: LoRA training, bigger batches, multi-ControlNet pipelines, fewer memory surprises
    • Typical comfort: “do the same things as B, but with headroom”
    • Pick C when: you want training + generation on one machine without dancing around VRAM
  • D — A100 80GB class (serious workloads)
    • Best for: large-scale training, heavy pipelines, big VRAM jobs, multi-task workflows
    • Pick D when: you already know what you’re doing — or you absolutely hate babysitting memory

Pick by task (most common scenarios)

Scenario 1 — SD 1.5 image generation (most users)

If your goal is “click & draw” (A1111, anime, portraits, character art), A is usually the best deal.

  • Start with: A
  • Upgrade to B if: you want bigger resolution / more ControlNet / smoother hires without compromises

Scenario 2 — SDXL generation

SDXL tends to want more VRAM for a comfortable experience. If you plan to live in SDXL, B is the “default comfortable pick”.

  • Start with: B
  • Consider C/D if: you’re also training, batching, or stacking many controls

Scenario 3 — Heavy ControlNet stacks (pose + depth + face + line)

ControlNet is additive memory pressure. The more you stack, the more you want headroom.

  • 1–2 ControlNet units → A is fine
  • 2–3 units + higher-res → B feels better
  • 3+ units + hires + “don’t want to think” → C/D

Scenario 4 — Training (LoRA / DreamBooth)

Training is where VRAM stops being a “nice-to-have” and becomes the main limiter.

  • Light LoRA experiments → B can work (with sane settings)
  • Regular LoRA training / fewer limits → C is the sweet spot
  • Bigger models, higher resolution, larger batch, faster iteration → D
Training warning: If you’re choosing A for training “to save money,” you may lose that money in time (slower iteration + more babysitting + more failures). For training-heavy users, C often ends up cheaper in real life.

Scenario 5 — High-res outputs (the “hires fix addict”)

If your workflow is “base → hires → polish” and you hate mush, you want headroom.

  • Hires ~1.3× and careful denoise → A works
  • Hires bigger, SDXL, heavier detail pipelines → B recommended
  • Large outputs + heavy controls + stability → C/D

VRAM rules of thumb (don’t OOM)

If you only remember one thing: memory pressure stacks. The biggest multipliers are:

  • Resolution (base width/height)
  • Batch size
  • Number of ControlNet units
  • Hires scale + denoise
  • Model family (SDXL generally heavier than SD 1.5)
Danger hint: If you hit OOM, reduce in this order: batch → hires scale → ControlNet count → base resolution. Don’t instantly blame the model. It’s usually one of these.

How to save money without suffering

  • Do ideation on A: compose, prompt, rough pose, quick drafts.
  • Upgrade only for the “final pass”: switch to B/C/D when you’re polishing or training.
  • Don’t pay for idle time: plan your prompt + reference first, then spin up compute.
  • Use the ladder: smaller hires steps are often better than one giant upscale.

FAQ

“I’m new. Which should I pick?”

Start with A. It’s the easiest and most cost-efficient way to learn. If you outgrow it, you’ll know exactly why (SDXL / hires / more ControlNet / training).

“Why do I OOM even though the same settings worked earlier?”

Memory pressure can change based on: model/checkpoint, ControlNet stack, browser tabs/apps, batch size, and whether hires/training is active. If it becomes unstable, reduce one multiplier at a time.

“Which group gives the best quality?”

Quality mostly comes from model + prompt + settings. Bigger GPUs don’t magically “draw better,” they just let you run heavier workflows (higher res, more controls, bigger training) without compromise.

What’s next?

Pick the machine that matches your workflow — not your pride.
← Back to SD Home