Up to 40% faster bare-metal RTX 5090s — built to stay up.
On-demand bare-metal 8× RTX 5090 nodes for training, inference and rendering — P2P-accelerated, in our own Tier III data center, SSH-ready in under 30 seconds.
$0.80 / GPU · hour — no egress fees
$ entrim launch --node 8x-rtx5090
✓ provisioning node-93 (8× RTX 5090, 256GB)…
✓ node ready in 28s — ssh root@node-93.entrim.io
$ nvidia-smi --query-gpu=name,memory.total --format=csv,noheader
8× NVIDIA GeForce RTX 5090, 32768 MiB8×
RTX 5090 / node
256 GB
GDDR7 / node
<30s
Cold-boot to SSH
99.9%
Uptime target
Performance
Up to 40% faster multi-GPU training & inference.
Consumer RTX cards ship with peer-to-peer transfers disabled, so multi-GPU jobs bounce every tensor through system memory. We run patched open GPU kernel modules that re-enable direct GPU-to-GPU P2P over PCIe — activations, weights and gradients move card-to-card without the detour, speeding up both distributed training and tensor-parallel inference.
- Direct GPU↔GPU transfers — no host-memory bounce
- Higher NCCL bandwidth for training and inference
- All 8 GPUs in the node pull together
Representative multi-GPU scaling — relative tokens/sec on the same hardware and job. Figures illustrative.
Thermals
Built to run flat-out — cool and quiet.
Eight RTX 5090s in one box throw a lot of heat. Each node lives in a custom-built chassis with a high-static-pressure fan wall and tuned front-to-back airflow, so every card holds its boost clocks through the longest runs — no throttling, no surprises.
- High static-pressure fan wall, N+1 redundant
- Tuned front-to-back airflow per node
- Sustained boost clocks — no thermal throttling
- 24/7 temperature & fan-health telemetry
The hardware
NVIDIA GeForce RTX 5090
The fastest consumer GPU ever built, on the Blackwell architecture. Every Entrim node packs eight of them — 256 GB of VRAM in one box — to fit larger models and bigger batches.
Every node ships with a full AMD Turin host.
The eight GPUs sit on a serious platform — plenty of CPU cores, memory and fast local storage to keep them fed — all on redundant power.
Reliability & compliance
Owned EU hardware, built to stay online.
Not a reseller. We run our own bare-metal in our certified EU data center — redundancy at every layer, and engineers who actually answer.
Our own Tier III data center
A concurrently maintainable Tier III data center in Slovenia, central EU — owned and operated by us, with redundant power and cooling paths and 24/7 on-site staff. No reselling.
ISO 27001 + GDPR
Information security managed to ISO 27001. Data is processed and stored in the EU under a fully GDPR-compliant chain — and wiped on instance release.
Dedicated public IP
Every instance gets its own dedicated public IPv4. Host an endpoint, whitelist by IP, open ports — no shared NAT, no noisy neighbours.
Bare-metal access
Full root on real bare-metal. Load custom kernel modules (including our P2P drivers), tune the OS and run any container — nothing is locked down.
Hot-spare swaps
If a node ever fails, we don't wait on an RMA — we migrate you to a pre-provisioned hot spare from the same rack and get you running again fast.
Redundant power
Dual hot-swap PSUs per node, UPS battery backup and on-site generators — your job rides straight through a utility power event.
RAID-protected drives
Local NVMe runs in redundant RAID, so a single drive failure never takes your data — or your training run — down with it.
Redundant switches
Every node has dual uplinks into redundant top-of-rack switches. No single switch, port or cable is a point of failure.
24/7 human support
Real engineers monitor the fleet around the clock. When you need help, you talk to the people who actually run the hardware.
Pricing
One node. One price.
Rent a full 8× RTX 5090 node, billed per second. No egress fees — stop an instance and billing stops with it.
≈ $6.40 / hour for the full 8-GPU node
- 8× RTX 5090 · 256 GB GDDR7 total
- AMD Turin · 64 cores · 384 GB DDR5
- 8 TB NVMe RAID · redundant power
- Bare-metal with full root access
- Dedicated public IPv4
- P2P drivers enabled — up to 40% faster
- Per-second billing · no egress fees
Need several nodes or a multi-node cluster? Contact sales for volume & reserved pricing.
FAQ
Questions, answered.
Spin up your first 8× RTX 5090 node.
Create an account, open the console and reserve a node. You only pay when an instance is running.