Hardware & Network

Supported GPUs

Your GPU model must appear on the Pricing catalog. Meshive’s operations team curates the supported list in the platform database, so anything not on that page is rejected at registration.

Tip

If your GPU is missing from the catalog but you would like to host it, contact support before you purchase. New models are added on a case-by-case basis after driver and stability validation.

AMD GPUs aren’t on board yet — they’re on the roadmap, stay tuned.

Same-model rule

A single machine must carry exactly one GPU model, with the same memory variant, on every slot.

Example configuration	Allowed?
8 × RTX 5090	✅
4 × B200	✅
4 × RTX 5090 + 4 × RTX 4090	❌ different model
2 × H100 80GB + 2 × H100 94GB	❌ different memory variant

This keeps pricing, scheduling, and per-GPU performance uniform for clients — and it is enforced automatically during registration.

PCIe generation and lanes

The host system (motherboard + CPU + slot configuration) must support each GPU at its maximum PCIe generation and lane width. Running a card on a slower link is rejected.

Concrete examples:

• ✅ RTX 5060 at PCIe 5.0 x8 — the card’s max generation and max lane width.
• ❌ RTX 5090 at PCIe 4.0 x16, PCIe 5.0 x8, or PCIe 5.0 x4 — the 5090 tops out at PCIe 5.0 x16, so anything narrower or slower is a no-go.

Public IP

The four ports below must be reachable from the public internet on a publicly routable IP. The cleanest configuration is to give the host machine its own public IP.

Forwarding just those four ports from a router with NAT is technically possible, but not recommended.

Network ports

Port	Protocol	Purpose
22	TCP	Host SSH (password is rotated monthly by Meshive)
2222	TCP	Pod SSH proxy — clients SSH into their pods through this port
443	TCP	Agent ↔ control-plane HTTPS
41641	UDP	Tailscale VPN

Recommended companion specs

Per-pod policy: min = 1 vCPU per 4 GB VRAM and 1 GB RAM per 1 GB VRAM, max = 2× min. A host should provision at least the max (plus a small OS/agent reserve) so any rentable pod fits.

Enter your per-GPU VRAM and GPU count to size the host:

Per-GPU VRAM (GB) GPU count

Per-pod	vCPU	RAM
Pod minimum	8	32 GB
Pod maximum	16	64 GB
Host recommended (≥)	16	64 GB

Add a small OS/agent reserve on top of the host-recommended numbers (a few vCPU and a few GB of RAM is plenty).

Storage configuration

Storage is the one resource clients can accumulate on a host (each pod’s persistent volume adds up), so more is generally better. Configuration differs between Compute Nodes and Storage Nodes — see below.

Compute Node

Three valid scenarios depending on how many SSDs the host has:

• 1 SSD (OS + data on the same disk). Partition the OS at 48–64 GB and leave the rest as raw free space. The agent will lay out a data partition in that free space.

• 2 SSDs (OS / data physically separated). Use the entire OS disk as the OS partition. Leave the data disk completely free. The agent picks the disk with the most free space for data — so as long as the data disk has more free space than the OS disk (which is full), this works automatically.

• 3 or more SSDs. ⚠️ Compute Nodes do not auto-assemble multiple disks into RAID. The agent still only picks the single disk with the most free space for the data partition. If you want multi-disk RAID on a Compute Node, build it yourself during OS install (LVM, mdadm, ZFS) and present the resulting volume as one disk.

See Install Ubuntu OS → Free space for the partition-layout walkthrough.

Storage Node

• Disk count: minimum 2, recommended 4 or more, and even-numbered to minimize capacity loss.
• Same capacity across all data disks — required.
• No host-side RAID — the Meshive agent assembles RAID for you automatically. Just leave every data disk as raw free space.
• Filesystem: handled automatically by the agent (XFS).
• Per-disk capacity, NVMe vs SATA: no specific requirement.