Customer guide · Accounts & compute

Create your account, rent compute, watch it go live.

You rent GPU capacity from a provider who runs Rackify. This page covers getting your account, renting machines by GPU model right in your console, and where to watch every GPU you’re paying for — live, isolated, and honest.

Step 1 · Your account

Redeem your invite.

Get an invite from your provider

Accounts are invite-only: your provider mints a single-use invite code bound to your company’s tenant (valid 14 days, shown to them once). Ask your provider contact for one per teammate who needs console access.

Register

  1. Open /register on your provider’s console URL.
  2. Enter your email, a password of at least 12 characters, and the invite code.
  3. Click Create account — you land in your console at /console, already scoped to your company’s tenant.

Account facts: passwords are argon2id-hashed; sessions last 30 days (use Log out on shared machines). Locked out? There is no “email me a reset link” — no email is ever sent by the platform — so ask your provider for a one-time reset code: it is single-use, expires in an hour, and you redeem it at /reset with your new password. Your account, tenant, and history are untouched; every other session of yours is signed out.

Your profile

The user menu in the top corner of your console opens Manage profile (/profile):

  • Display name and photo — what your teammates see next to your account. Photos are PNG, JPEG, or WebP, resized small before upload. They are never public and never on a CDN: the image is served only to signed-in people on your own tenant and to your provider’s operator accounts. Anyone else — including another provider’s customers — gets the same “not found” as an account with no photo at all.
  • Email — changing it requires your current password, and the new address becomes your login immediately.
  • Password — also gated on your current password; changing it signs out every other session you have, and keeps the one you’re using.
  • Show simulated data — off by default. Demo machines your provider runs for evaluation are hidden from your dashboards unless you switch this on; whenever something is hidden, the page says how many, so an empty view is never ambiguous.

Step 2 · Rent compute

Rent GPUs right in your console.

Renting is self-service: your console shows live availability by GPU model from your provider’s own capacity pool — their machines, never a shared marketplace, and never another provider’s hardware. Rent whole nodes and they’re assigned to your tenant immediately; release them just as fast when you’re done.

Check availability, pick a model, rent

How renting works
  • Availability by GPU model — nodes available, GPUs per node, and total GPUs, counting only machines online right now
  • Whole nodes — full servers (typically 8 GPUs each), 1–8 nodes per request; MIG slices are on the roadmap
  • Rent — the machines are assigned to your tenant on the spot
What happens next
  • Each machine appears in your console immediately
  • Telemetry starts from the moment of assignment — utilization, VRAM, thermals, power, faults at 15-second resolution
  • Partial fulfillment is reported honestly: you see exactly how many nodes were assigned and any shortfall — never a silent gap
  • You verify what you got: models, VRAM, MIG mode, and interconnect topology are all inspectable per node

Verify before you pay for it: the console shows each GPU’s UUID, model, VRAM, and MIG mode, plus lifetime ECC and Xid fault counters — a MIG slice is never presented as a full GPU, and unhealthy silicon shows up in the numbers.

Scaling up, down, or moving off

More capacity is the same loop — check availability, rent. Done with a machine? Release it from your console: it leaves your fleet immediately and returns to your provider’s pool, where only their customers can rent it, and your historical telemetry stays yours (history is never rewritten). Data residency follows your provider’s deployment: on sovereign deployments, your telemetry and logs live in their dedicated cloud account and region — never a shared control plane.

When to go through your provider

Your provider contact is still the path for anything the pool doesn’t answer: GPU models not showing in availability, reserved or committed terms, MIG slices, or site/region constraints. They can also assign machines to your tenant directly — those appear in your console the same way.

Coming: more self-service

The sidebar previews what’s rolling out next: Instances (launch on-demand, reserved, and capacity blocks as workloads), Notebooks (JupyterLab on your rented GPUs), Inference (an OpenAI-compatible API on managed model endpoints), and Billing (invoices, usage detail, and payment methods). Each appears today as a visible SOON item so you always know what’s product and what’s roadmap.

Step 3 · Your console

Watch every GPU you pay for.

Everything your console shows you, and why you can trust it.

Read the Overview

  • Tiles — your total GPUs, nodes online ( n of m), live power draw across your fleet, and open health conditions.
  • GPUs by model — live availability: how many of each model are online right now.
  • Utilization & power charts — your whole fleet over 24h / 7d / 30d. Every chart has a table-view toggle (also great for copy-paste).
  • Your nodes — each machine with site, GPU count, status, and last-seen time.

Data refreshes every 15 seconds. Online means the machine has heartbeated within the last 2 minutes.

If you see a SIMULATED badge, you’re looking at a demo environment — synthetic machines your provider uses for evaluation. Real hardware never carries the badge.

Drill into a machine

Click any node to open its page: full GPU inventory (model, VRAM, MIG mode, UUID), live per-GPU charts for utilization, VRAM, temperature and power, plus lifetime ECC and Xid fault counters — the numbers that tell you whether the silicon you’re paying for is healthy.

Tail your machines’ logs

The Logs panel on each node page shows your machine’s GPU-relevant logs — driver messages, dmesg lines, agent events — with Xid faults highlighted. Toggle Follow for a live 3-second tail; filter by level or source. It’s your hardware, so logs are visible to you by default (providers can disable this per contract).

Trust what you’re reading

  • Isolation, at both levels: your console can only ever return your own tenant’s data, and your provider’s console can only ever return their own datacenter’s. Both scopes are enforced by the platform on every query — derived from the session, never from anything the browser sends — and each is verified by its own automated leak test on every release.
  • Honest gaps: if a machine loses network, its agent buffers locally and backfills history on reconnect (up to a 55-minute replay window). Charts never interpolate or smooth: a gap is shown as a gap, whether the data never existed or couldn’t be backfilled.
  • Offline flags: a machine marked offline has actually missed 2+ minutes of heartbeats — it’s computed from the same data you see.