Switch GPUs Without Losing Your Setup — How Disk Retention Works

Switching GPU types on a cloud instance usually means rebuilding from scratch — re-downloading models, reinstalling packages, reconfiguring tools. On SynpixCloud, your entire environment persists across GPU changes through disk retention. Keep an RTX 3090 workspace, attach RTX 4090 compute to it for a batch run, then scale back down. No re-setup. Retention costs $0.0001/GB/hr while the disk is preserved.

Say you have 80 GB of models, a tuned ComfyUI environment, and three days of LoRA training outputs on an RTX 3090. Now you need an RTX 4090 for a production batch. On most platforms, you either re-download everything on a new machine, or keep paying for the old one while the new one sets up. Both options waste time and money.

There is a third option: keep the disk, switch the GPU.

The Rebuild Penalty Is Real

Every time you move between GPU types on a traditional cloud platform, you rebuild from scratch. The environment does not follow you.

For a typical AI workflow, that means:

That is 1-3 hours of setup time that produces zero useful output. And if you switch GPUs frequently — scaling up for batch jobs, scaling down for interactive work — you pay this penalty every single time.

For context on how model sizes are growing: see our SDXL VRAM requirements guide. The rebuild penalty only gets worse as models get larger.

Retained Disk: Your Environment as a Portable Asset

The core idea is simple: separate compute from state.

Your GPU is temporary. Your environment — the models, code, configurations, training outputs — is durable. When you retain your disk, you are not just saving files. You are preserving a complete, ready-to-run workspace that can attach to any compatible GPU.

Think of it as saving a game. The next time you load, everything is exactly where you left it — on whatever hardware you choose to run it on.

What gets preserved:

• All installed models (checkpoints, LoRAs, VAEs, ControlNets)
• Python environments and pip/conda packages
• Custom nodes and their dependencies
• Project outputs, logs, and intermediate results
• Configuration files and environment settings
• Training checkpoints and fine-tuned weights

What changes:

• The GPU attached to it
• The billing period
• Optionally, the compute configuration

GPU Switching: How It Works

When your instance expires or you stop it with disk retention enabled, the GPU is released but your disk stays intact. From this retained disk, you can create a new instance with a different GPU type.

The process:

1. Your instance reaches end of life — billing period ends, or you stop it manually
2. Disk enters retention — compute stops, retention billing begins ($0.0001/GB/hr)
3. You recreate from disk — choose a new GPU type, billing period, and launch
4. Boot with full environment — your workspace is exactly as you left it

Most environments recreate successfully across GPU types. CUDA applications, Python environments, and model files are generally GPU-agnostic at the software level — the same PyTorch code runs on an RTX 3090 and an RTX 4090 without modification. The underlying drivers and CUDA runtime handle hardware abstraction.

In rare cases involving highly specialized kernel configurations or hardware-specific compiled extensions, minor adjustments may be needed. But for the vast majority of AI workflows — Stable Diffusion, ComfyUI, LLM inference, LoRA training — switching GPU types is seamless.

Four Scenarios Where This Matters

1. Scaling Up for Production

You develop and iterate on an RTX 3090 at $0.30/hr. When the workflow is ready for a production batch — hundreds of images, a full training run, video generation — you switch to an RTX 4090 at $0.39/hr for roughly 2x the throughput. No re-setup. The batch runs at full speed from minute one.

When the batch is done, you scale back down. Same disk, lower cost.

2. Testing Before Committing to Hardware

Considering buying an RTX 4090 to replace your local RTX 3090? Instead of spending $1,400+ on hardware, rent a cloud RTX 4090 for a week with your exact environment. Run your real workloads. Measure the actual speedup on your specific pipelines — not synthetic benchmarks.

If the 2x speed saves you meaningful time on billable work, the upgrade pays for itself. If your work is not speed-sensitive, you just saved $1,400.

Use the cost calculator to estimate what a test period would cost.

3. Budget Optimization

Not every task needs the fastest GPU. Training a LoRA? An RTX 3090 handles it fine. Running inference on a trained model at scale? The RTX 4090's throughput advantage makes it more cost-effective per output.

With retained disk, you can match GPU power to the task:

Compare GPU specs and pricing on the GPU comparison tool, or see the full cloud GPU pricing comparison to find the right fit for each phase of your workflow.

4. Availability-Based Flexibility

GPU availability fluctuates. If your preferred RTX 4090 is temporarily out of stock, you can start on an available GPU — RTX 3090, A5000, or even Tesla V100 — and migrate to the 4090 when it becomes available. Your work does not stall waiting for a specific GPU type.

Browse current availability on the marketplace.

The Economics of GPU Switching

Let's compare three approaches for a user who needs both development time (20 hrs/week on RTX 3090) and production runs (5 hrs/week on RTX 4090).

Approach A: Two separate instances (traditional)

Approach B: Single RTX 4090 for everything

Approach C: Retained disk with GPU switching

Approach C saves $0.64/week over A (no rebuild penalty) and $1.76/week over B (right-sized GPU per task). Over a month, that is $3-7 in savings — plus the 45-60 minutes of setup time you never spend.

See current RTX 3090 and RTX 4090 pricing on the marketplace.

For high-VRAM workflows, the savings scale further. An A100 at $0.63/hr for development time you could handle on a $0.30/hr RTX 3090 wastes $0.33/hr — $6.60 per week.

Compatibility Across GPU Types

Most AI software is designed to be hardware-portable. Here is what to expect when switching between common GPU types:

The general rule: if your software runs on "CUDA" rather than on a specific GPU model, it will work across GPU types without changes.

For an overview of which GPUs fit which workloads, see our guide on choosing the right GPU.

Retain Disk Is Not Just Storage

The traditional mental model of "cloud storage" is a hard drive you rent. You put files on it. You take files off it.

Retained disk is different. It is an environment persistence layer:

• It preserves not just files, but installed software, configured services, and runtime state
• It turns compute into a swappable component while the workspace stays fixed
• It makes GPU selection a per-task decision rather than a per-setup decision

This reframes how cloud GPU workflows operate. Instead of "I need to set up an RTX 4090 instance," the question becomes "I need RTX 4090 compute attached to my workspace." The workspace is the constant. The GPU is the variable.

For a deeper dive into why environment persistence matters, read why persistent disks matter more than GPU specs.

Can I Switch GPU Types Without Rebuilding?

Yes. Enable disk retention on your instance. When you stop or let it expire, the disk preserves your full environment — models, packages, configs, outputs. You then recreate the instance from that disk with a different GPU type. Your workspace boots exactly as you left it, on the new hardware. Most AI software (PyTorch, ComfyUI, Stable Diffusion, TensorFlow) is GPU-agnostic and works across GPU types without modification. Retention costs $0.0001/GB/hr while the disk is preserved.

Getting Started

1. Create an instance on the marketplace with disk retention enabled
2. Set up your environment — install models, configure tools, run your workflows
3. When you need a different GPU, let the instance expire or stop it (disk is retained automatically)
4. Recreate from disk — choose a new GPU type and billing period
5. Resume immediately — everything is where you left it

Disk retention costs $0.0001/GB/hr — about $0.48/day for a standard 200 GB disk. That is the cost of keeping your entire workspace portable across any GPU type we offer.

---

Your environment is the product of hours of setup. Your GPU is the tool you attach to it. On SynpixCloud, you can change the tool without rebuilding the product. Browse available GPUs and start with disk retention enabled.

---

Related guides:

• RTX 3090 vs RTX 4090 for AI — Head-to-head comparison with cloud rental economics
• Cloud GPU vs Local GPU: Real Cost Analysis — When to buy vs rent
• Cloud GPU Pricing Comparison 2026 — Full provider breakdown
• Why Persistent Disks Matter More Than GPU Specs — Deep dive into environment persistence
• RTX 4090 vs A100 vs H100 Comparison — When to scale up to data center GPUs