Choose a machine based on the work
Move macOS work to a dedicated physical machine: build a release, connect to your team’s pipeline, access a remote development desktop, or test MLX inference. Two Apple Silicon M4 configurations, never virtualized, from $20.2/day.
SSH and VNC access · Billed in USD · Availability is shown live in the console
Independent iOS developers
If you don’t have a Mac locally—or your main machine is busy—sync your project to the cloud for builds, simulator regression tests, and archiving. MetalBench 16 is a good starting point for a single app’s build pipeline. Consider MetalBench 24 when you need desktop tools and multiple tasks running at once.
Prepare project dependencies and a legitimately obtained signing certificate and provisioning profile, then verify your command-line build and signing workflow. You’re responsible for meeting the permissions and compliance requirements for App Store submissions. Confirm your existing automated upload method works before scheduling a release.
A cloud physical Mac doesn’t include a remotely controlled iPhone. Run simulator tests on the machine, but use an environment with the relevant device for debugging that requires USB peripherals, phone sensors, Bluetooth, or touch input. Remote desktop does not pass through peripherals.
Daily billing can suit a short task, but buying day passes repeatedly may not make sense for a release week. Estimate how long you’ll need the machine, then compare daily, weekly, monthly, and quarterly prices. If a short task runs longer than expected, check your renewal options.
Rent MetalBench 16Record the Xcode version, dependency manager, and project branch. Run an initial build to rule out toolchain differences.
Run simulator tests and save the test report, failure screenshots, and repeatable build command.
Export the archive, symbol files, and logs. Back up the deliverables to storage you manage, then end the session.
Acceptance check: you can regenerate the deliverables from the same commit—not just get one successful build on the desktop.
CI/CD teams
Register the machine as a GitHub Actions or GitLab self-hosted runner and configure the toolchain over SSH. Use labels to route projects and jobs. CPU, memory, and storage on the physical machine aren’t shared with other tenants, but you’ll need to manage concurrency on the same machine.
For everyday multi-target projects, start by evaluating MetalBench 24: M4, 24GB memory, and a 512GB SSD. Measure peak memory, disk usage, and the time for a complete build before deciding how many runners to launch. More concurrent jobs don’t automatically mean faster delivery.
Manage build caches, source workspaces, and final deliverables separately. Remove old simulators and derived data before caches fill the disk. Need more capacity? Choose a +1TB SSD or +2TB SSD when ordering. Check the console for available configurations.
For a node you’ll use regularly, MetalBench 24 costs $532.8/quarter, less than buying three consecutive months. Choose a billing period that matches your project; don’t let a short-term price comparison obscure the full upfront payment.
Rent MetalBench 24Register a runner and restrict access to the repositories or projects that need it. Use labels to keep jobs out of other environments.
Specify the Xcode path and dependency versions in each job. Run xcodebuild tests first, then add parallel jobs.
After each job, remove temporary keys, certificate copies, and tokens. Keep execution logs free of sensitive information.
For multi-machine workflows, consider Thunderbolt 5 clustering. You’ll still need to validate the application protocol, topology, and data-sync method separately; clustering does not combine memory across machines.
Remote work across borders
Choose from four Asia-Pacific locations: Singapore, Tokyo, Seoul, and Hong Kong. Use VNC for the macOS desktop and SSH for builds and file management. Test your network connection first, then decide how much daily work to move to the cloud.
Use VNC to view the interface, check layouts, and investigate graphics issues. For long builds, batch processing, and continuous logs, SSH is usually the better fit. Don’t rely on a single foreground terminal if your connection drops briefly: use a session manager and save output to files as you go.
For everyday development, start by evaluating MetalBench 24 at $195.9/month. If you only need code editing and light builds, try MetalBench 16 first to validate your workflow. Don’t upgrade before measuring memory pressure.
Remote desktop works well for development tools and admin interfaces, but shouldn’t be assumed to feel like a local monitor. Test first if you need high frame rates, accurate color, or local peripherals. On a weak connection, lower the resolution and reduce animations and unnecessary windows.
View SSH and VNC setup guideChoosing a location is a network-routing decision, not a performance guarantee. Your experience also depends on your local broadband, routing, packet loss, and client settings.
AI experimenters
Use MLX on M4 to test inference with quantized models. Apple Silicon’s unified memory architecture reduces data transfers between CPU and GPU, but model weights, runtime, and context all use the same memory. More storage doesn’t make up for less memory.
Start by evaluating MetalBench 24, with an optional +2TB SSD for models, datasets, and experiment output. Quantized 7B–13B models are worth testing, but aren’t guaranteed to run: resource needs vary with quantization, context length, batch size, and implementation.
Use a fixed prompt to measure time to first output, generation speed, and peak memory, then increase the context length. If memory pressure persists, reduce the context, batch size, or model size instead of masking the bottleneck with more background processes.
For a short proof of concept, choose a short billing period. Consider quarterly billing only after confirming you can reproduce the environment. Don’t present inference experiments as large-scale training resources; if a task exceeds the machine’s memory budget, reassess the model and execution plan.
Configure MetalBench 24+2TB SSD is an optional add-on:$4.8/day · $13/week · $24.2/month · $65.8/quarter, not included in the base machine price.
Match the machine to the task
These are starting recommendations, not guarantees of concurrency or model performance. Run a representative task first, then adjust based on peak memory, workspace size, and how long you need the machine. All machines are dedicated physical machines.
| Task | Recommended configuration | Starting period | Why this configuration |
|---|---|---|---|
| Light builds | MetalBench 16 M4 / 16GB / 256GB | Daily $20.2/day | Start with a single-app CI build and simulator regression tests. Check dependency installation, build caches, and available disk space. |
| Everyday development | MetalBench 24 M4 / 24GB / 512GB | Monthly $195.9/month | More room for multi-target projects, desktop tools, and parallel simulators. Validate concurrency with your own workloads. |
| Model inference experiments | MetalBench 24 Optional +2TB SSD | Quarterly after validation | Use extra storage for models and output. The base machine and add-on are billed separately; first confirm there’s enough memory for your workload. |
| Multi-machine workflows | Multiple MetalBench 24 machines Thunderbolt 5 clustering available on request | Quarterly after settling into regular use | Split workloads across separate machines; each is billed independently. Clustering is charged per machine as an add-on. Confirm topology and software support before deployment. |
Thunderbolt 5 clustering add-on, per machine: $1.1/day · $3/week · $5.5/month · $15/quarter. Confirm the final machine, location, billing period, and add-ons when ordering.
Choose how you’ll connect, too
All four locations support the workflows below. Recommended locations are starting points, not latency guarantees; test the network route from your own connection. Export your deliverables before ending each task, and back up your business data independently.
Sync the source → restore dependencies → run simulator regression tests → export signed deliverables. Keep failure logs so you can continue troubleshooting locally.
Register a runner → pin the toolchain → control concurrency → upload test reports. Caches can be rebuilt; save final deliverables separately.
Test the connection → adjust image quality → sync work files → lock the screen when you leave. Handle interactive work separately from long-running background tasks.
Prepare a quantized model → fix the inputs → measure memory and speed → save the output. Confirm it runs before extending the experiment.
Payment is accepted only via USDT-TRC20 or Visa, Mastercard, and Amex through Stripe. All prices are in US dollars (USD); available payment gateways are confirmed by the backend. If you have trouble connecting or running a task, sign in to the console to submit a support ticket.
Prepare your project and acceptance criteria
Run a real task first, then decide on a long-term configuration. Not sure what you need? Tell us about your project size, expected usage period, and access location, and we’ll help you check the resource requirements.