Infrastructure

# What we rack, and what happens when it breaks.

A dedicated server is a promise about a physical object. This page describes that object: which cards we buy, how a node is tested before it takes a tenant, and what we do the day one of them dies.

- 12 GPU models we operate
- 72 h burn-in before a node takes a tenant
- 4 h target to replace failed hardware
- 0 cards resold from another provider

## Every card we operate

This is the whole list — there is no unlisted tier and no card we keep for larger accounts. Memory bandwidth is on the table because for inference it predicts throughput far better than any teraflop figure does.

**GPU models, architecture, memory, bandwidth, node sizes and entry price**

| Card | Architecture | Memory | Bandwidth | Node sizes | From |
|---|---|---|---|---|---|
| L4 PCIe add-in card | Ada Lovelace | 24 GB GDDR6 | 300 GB/s | 1×, 2×, 4×, 8×, 10× | $125/mo |
| RTX 4090 PCIe add-in card | Ada Lovelace | 24 GB GDDR6X | 1,008 GB/s | 1×, 2×, 4×, 8× | $193/mo |
| RTX 5090 PCIe add-in card | Blackwell | 32 GB GDDR7 | 1,792 GB/s | 1×, 2×, 4×, 8× | $335/mo |
| RTX A6000 PCIe add-in card | Ampere | 48 GB GDDR6 ECC | 768 GB/s | 1×, 2×, 4×, 8× | $286/mo |
| L40S PCIe add-in card | Ada Lovelace | 48 GB GDDR6 ECC | 864 GB/s | 1×, 2×, 4×, 8× | $714/mo |
| A100 PCIe PCIe add-in card | Ampere | 40 GB HBM2 | 1,555 GB/s | 1×, 2×, 4×, 8× | $392/mo |
| A100 PCIe PCIe add-in card | Ampere | 80 GB HBM2e | 1,935 GB/s | 1×, 2×, 4×, 8× | $1,091/mo |
| H100 PCIe PCIe add-in card | Hopper | 80 GB HBM3 | 2,000 GB/s | 1×, 2×, 4×, 8× | $1,469/mo |
| A100 SXM4 SXM module, HGX board | Ampere | 80 GB HBM2e | 2,039 GB/s | 4×, 8× | $4,395/mo |
| H100 SXM5 SXM module, HGX board | Hopper | 80 GB HBM3 | 3,350 GB/s | 4×, 8× | $6,061/mo |
| H200 SXM5 SXM module, HGX board | Hopper | 141 GB HBM3e | 4,800 GB/s | 4×, 8× | $8,405/mo |
| B200 SXM6 SXM module, HGX board | Blackwell | 180 GB HBM3e | 8,000 GB/s | 4×, 8× | $11,790/mo |

Every card is a retail or OEM part bought new, with its warranty in our name. We do not buy ex-mining stock, and we do not buy cards whose history we cannot account for — a used RTX 4090 is cheap for a reason, and the reason arrives in month four.

## What surrounds the cards

A GPU starved of CPU, RAM or disk throughput is an expensive way to wait. The chassis is chosen for the card count, not sold as an upgrade path.

**Chassis specification by GPU count**

| Node | CPU | System RAM | Local storage | Port |
|---|---|---|---|---|
| 1 × GPU | AMD Ryzen 9 7950X 16 c / 32 t · or Intel Xeon Silver 4410Y | 128 GB | 2 × 2 TB NVMeGen4 U.2, no RAID imposed | 1 Gbit/s |
| 2 × GPU | AMD Threadripper 7960X 24 c / 48 t · or Intel Xeon Gold 5416S | 256 GB | 2 × 3.84 TB NVMeGen4 U.2, no RAID imposed | 2 Gbit/s |
| 4 × GPU | AMD Threadripper 7970X 32 c / 64 t · or 2 × Intel Xeon Gold 6438Y+ | 512 GB | 4 × 3.84 TB NVMeGen4 U.2, no RAID imposed | 10 Gbit/s |
| 8 × GPU | 2 × Intel Xeon Gold 6438Y+ 64 c · or 2 × Intel Xeon Platinum 8462Y+ | 1 TB | 8 × 3.84 TB NVMeGen4 U.2, no RAID imposed | 25 Gbit/s |
| 10 × GPU | 2 × Intel Xeon Platinum 8462Y+ 64 c · or 2 × Intel Xeon Platinum 8462Y+ | 1 TB | 8 × 7.68 TB NVMeGen4 U.2, no RAID imposed | 25 Gbit/s |

Where two CPUs are listed, either may be delivered — they are matched on core count and memory channels, and the configurator shows you which one your node has before you pay. Disks are handed over raw: we impose no RAID level, because the right one depends on whether you are storing checkpoints or a dataset you can re-download.

## Before a node is sold

A machine reaches the catalogue only after it has spent three days failing to break. Almost every card that will die early dies here, in our hands, and not in yours.

STEP 1 · 2 HOURS

### Assembly and inventory

Serial numbers of every card, module and disk are recorded against the chassis. That record is what lets us tell you, later, exactly which physical part is in your machine.

STEP 2 · 24 HOURS

### Memory and compute soak

Full-VRAM ECC pattern tests on every card, then sustained matrix multiplication at 100% utilisation. A single uncorrectable error and the card goes back in the box.

STEP 3 · 24 HOURS

### Thermal and power soak

The whole node held at full load while inlet temperature is raised to the suite's worst case. We record the clock the cards actually sustain — not the boost figure on the box — and a node that throttles below spec is re-seated or re-pasted before it is tested again.

STEP 4 · 24 HOURS

### Interconnect and storage

NVLink or PCIe peer-to-peer bandwidth measured card to card, NVMe write endurance sampled, port held at line rate. The numbers are compared to the other nodes of the same model, because a node that is 15% slow is a node with a problem.

THEN · RACKED AND WAITING

### Available, with an image ready to write

This is why delivery is under 5 minutes and not a working day: nothing is assembled when you order. The machine already exists, powered and tested, and your payment triggers an image write, not a build.

## Firmware and drivers

Driver stack Installed, pinned, yours to change Every Linux image ships with a current NVIDIA driver, CUDA, cuDNN and NCCL already working. Nothing stops you replacing all of it — it is your root.

Firmware updates Never during your term BIOS, BMC and VBIOS are set before the node enters the catalogue. We do not push firmware to a machine someone is renting; it is applied between tenants.

Clocks Stock, and unlocked to you We ship every card at its reference clock and power limit. You may raise or lower them with `nvidia-smi`; the warranty consequence is ours, not yours.

Out-of-band access IPMI, on a separate network Remote power, serial console and virtual media, reachable even when the operating system is not. It is on a management VLAN that never touches your public port.

## When something fails

Hardware breaks. What distinguishes providers is not whether it happens but what the next four hours look like.

### Spares are on site

Each suite holds spare cards, PSUs, disks and one complete chassis per model. A replacement is a walk down an aisle, not a courier booking.

### Four-hour target

From your report to functioning hardware, around the clock. If we miss it, the SLA extends your term automatically — you do not have to claim it.

### Your disks stay yours

A failed GPU or PSU is swapped with the disks untouched and in place. We only ask before touching storage, and only when storage is what failed.

**A failed disk is not readable by the next tenant.** Drives removed from service are erased if they still respond, and physically destroyed if they do not. A dead disk cannot be wiped, so it is never returned to a manufacturer under warranty — we absorb that cost rather than let a drive that once held your data leave the building intact.

## Between two tenants

The same machine will be rented again after you. Everything below happens before it re-enters the catalogue, whether or not anyone asks.

**Decommissioning steps between two tenants**

| Step | What happens | Verified by |
|---|---|---|
| 1. Network | The machine is cut from its public port the moment the term ends. Nothing is left reachable while it waits. | Port state, logged |
| 2. Storage | Every NVMe device receives a full cryptographic erase, then a single-pass overwrite of the whole address space. | Read-back sample on each device |
| 3. GPU memory | Cards are reset and their memory cleared; ECC counters are read and compared to the values recorded at burn-in. | `nvidia-smi` reset log |
| 4. Firmware | BMC and BIOS settings are reflashed to our reference, discarding any change made during the term. | Checksum against reference |
| 5. Identity | IPMI credentials rotated, IP addresses returned to the pool, rDNS cleared. | Address held out until clean |

This is also why we cannot restore anything after a term ends. There is no grace period during which your data still exists somewhere — step 2 has already run. If you need it, take it off the machine before the term closes.

## What we will not do

**Partition a card.** No MIG, no vGPU, no time-slicing. One tenant per physical GPU, which is what makes the throughput on the listing reproducible.

**Oversubscribe a node.** There is no second account on the machine and no hypervisor between you and the metal.

**Resell someone else's capacity.** Every card on this page is a card we bought and racked. Nothing here is a reseller margin on another cloud.

**Substitute a different card.** If the model you ordered cannot be delivered, we do not quietly ship a near-equivalent — we tell you and refund the term.

**Enter your operating system.** Root is yours alone. We have IPMI, which can power-cycle and mount media; we do not hold a login inside your machine.

**Keep a copy of anything.** We take no backups of your disks. Snapshots exist only if you buy them, and they are deleted with the term.

## The specification on the row is the specification you get.

Every card above is available in all 6 data centres, racked and burned in, with an image ready to write.

[Browse the catalogue](https://gpuserver.io/#catalog) [Read the guides](https://gpuserver.io/guides)

## Related pages

- [Network Unmetered ports up to 25 Gbit/s, two carriers and an IX per site, always-on DDoS filtering, routed IPv6 — and the four things we do not offer, stated up front.](https://gpuserver.io/network)
- [Documentation First SSH, verifying the hardware, CUDA containers, serving a model with vLLM, RAID, firewalling, IPMI and reinstalls — the commands, on a real machine.](https://gpuserver.io/docs)
- [Service level agreement The 99.9% commitment: what counts as downtime, how it is measured from outside, five minutes of term back per minute lost, and the exclusions in full.](https://gpuserver.io/legal/sla)
- [About us Who runs gpuserver.io, why we buy the hardware instead of reselling it, and the four rules we do not bend — including the ones that cost us sales.](https://gpuserver.io/about)

---

Source: https://gpuserver.io/hardware/. This file is generated from the same data as the website; if a figure here differs from a page, the page is authoritative and this file is stale — the canonical source is https://gpuserver.io/.