Raspberry Pi 5 Home Cloud — 4 TB NVMe, always on

A compact, low-power home server: private cloud storage, remote SSH, video streaming, MacBook backups, Linux and C++ development, and moderate real-time processing — all from a board that draws about as much power as a phone charger. This note is the full plan, from which parts to buy to what the finished system is and is not good at.

In one line. A Raspberry Pi 5 (8 GB) with a 4 TB M.2 2280 NVMe SSD on a PCIe NVMe HAT, reached over Tailscale, costs €635–712 to build at August 2026 German prices and about €1–1.50 a month to run. Two thirds of that is the SSD. The drive is far faster than the network it is served over — and that is fine.

What this system is

The Pi 5 becomes a small always-on home server. One board, one SSD, one network cable, and a VPN that makes the whole thing reachable from anywhere without opening a single port to the public internet.

The hardware

Core parts

Optional extras

• Cat 6 Ethernet cable
• UPS, for safe shutdown during power outages
• A second storage device for backup
• USB4 / Thunderbolt NVMe enclosure, for direct MacBook use (§3.1)
• 2.5 GbE USB adapter, if higher network speed is required (§17)

Which NVMe SSD fits

The SSD type that matters is:

M.2 2280 NVMe
PCIe
M-Key

The number 2280 is the physical size, not a model name:

2280
││
│└── 80 mm long
└── 22 mm wide

The GeekPi N04 in this plan supports M.2 2230, 2242, 2260 and 2280. For a 4 TB SSD, 2280 is the format to buy — the choice at that capacity is far wider.

How the NVMe connects

The Pi 5 exposes a PCIe interface. The HAT is the physical adapter between the M.2 socket and that connector — nothing more clever than that, which is why it is reliable.

PCIe HAT vs USB enclosure

For a permanent home server, use the PCIe NVMe HAT. It is a cleaner installation, it frees the USB ports, it is a direct PCIe connection, it is well suited to 24/7 operation, and the Pi can boot Raspberry Pi OS directly from it (§6.2).

Use a USB enclosure instead only if you expect to disconnect the SSD often and move it between computers.

The same SSD and the MacBook

Direct connection over USB4

The same 4 TB NVMe SSD works with a MacBook Pro — but the GeekPi HAT cannot be used as a Mac adapter. For a direct connection you need an enclosure:

Why the SSD should stay on the Pi

For a home cloud, do not physically move the SSD back and forth. Keep it permanently attached to the Pi and reach it over the network:

MacBook Pro
     │
     │ Ethernet / Wi-Fi / Tailscale
     ▼
Raspberry Pi 5
     │
     ▼
4 TB NVMe

Any of these protocols will do it: SMB, SFTP, SSH, rsync, SCP, WebDAV, Jellyfin, Nextcloud, Syncthing.

Network performance — the real bottleneck

The Pi 5's Ethernet is Gigabit. That caps everything:

$$ 1 \text{Gbit/s} \approx 125 \text{MB/s}, \qquad \text{practical} \approx 100\text{–}115 \text{MB/s} $$

So when the Mac reads the SSD over the LAN:

NVMe SSD capability        ~400–500 MB/s on Pi
Gigabit Ethernet limit     ~100–115 MB/s

The network, not the SSD, is the bottleneck — by a factor of about four. Buying a faster SSD does not change this number; only a faster network does (§17).

Even so, 100 MB/s is more than enough for large document storage, photo libraries, backups, 1080p streaming, most 4K streaming, programming files, Git repositories, and remote development.

Remote access and file sharing

Tailscale from outside the home

The safest simple solution is Tailscale: an encrypted mesh that makes the Pi reachable as if it were on the local network, without exposing port 22 to the public internet.

ssh user@raspberrypi          # by hostname
ssh user@100.x.x.x            # by Tailscale IP

scp file.zip user@raspberrypi:/storage/
rsync -av ~/Documents/ user@raspberrypi:/storage/Documents/

SMB to macOS Finder

On the home network, Samba/SMB is the convenient option. Share /storage, then in Finder:

Finder → Go → Connect to Server
smb://raspberrypi.local/storage

The Pi storage then appears in Finder like any other network drive.

Storage layout and booting

Suggested layout

/storage
├── Documents
├── Photos
├── Videos
├── Music
├── Projects
├── Backups
├── Git
├── Docker
└── Shared

Raspberry Pi OS itself can live on the same NVMe drive:

4 TB NVMe
│
├── Raspberry Pi OS
│
├── /home
│
└── /storage
    ├── Documents
    ├── Photos
    ├── Videos
    ├── Projects
    └── Backups

Booting from NVMe

The Pi 5 can boot from NVMe once configured, so the microSD card is not needed for daily use. Either put the OS and the data on one filesystem, or split it:

4 TB NVMe
├── OS partition
└── Data partition

Video streaming

Jellyfin is the recommended media server: it reads files straight off the NVMe and serves them to the LAN or, through Tailscale, to anywhere.

Direct play

Direct play is the ideal case: the Pi reads the existing file and sends it to the client without converting it. This works very well for 1080p, 4K, 4K60 (given the bitrate and network capacity), H.264, and HEVC/H.265 where the client supports it.

Transcoding — the limit

Real-time high-end transcoding is far heavier, and the Pi 5 is not the right machine for:

• heavy 4K transcoding
• multiple simultaneous 4K transcodes
• very high-bitrate conversion
• complex AV1/HEVC conversion workloads

A stronger mini-PC or desktop is the answer for those.

Upload speed for remote viewing

Watching from outside the home, the limit is usually the home internet upload speed, not the Pi:

Video bitrate:   30 Mbps
Required upload: comfortably above 30 Mbps

A high-quality 4K stream can need tens of Mbps. The SSD is far faster than any of this.

Development on the Pi

C and C++ tooling

The Pi 5 is a very capable Linux C/C++ machine. The usual tools are all there:

gcc   g++   clang   cmake   make   ninja   gdb   git

#include <iostream>

int main()
{
    std::cout << "Hello from Raspberry Pi 5\n";
    return 0;
}

g++ -O3 main.cpp -o app     # -O2 / -O3 enable compiler optimization
./app

The code compiles to native 64-bit ARM instructions — nothing is emulated.

What the CPU is good for

The Pi 5 uses a quad-core 64-bit ARM Cortex-A76-class CPU.

What the NVMe actually speeds up

The SSD does not make CPU arithmetic faster. It improves everything that touches the filesystem: compilation, file loading, linking, large datasets, Docker images, databases, logging, video access, project indexing, package installation, swap behaviour, and general responsiveness.

C++ computation          File / data access
      │                        │
      ▼                        ▼
   Pi 5 CPU                4 TB NVMe

Real-time processing

The Pi 5 can do fast real-time work — but "real-time" depends entirely on the workload and the latency requirement.

What it handles well

Sensor processing, robotics control, audio DSP, network packet processing, moderate SDR workloads, telemetry, automation, and low-latency C++ applications. Standard audio rates are comfortable for well-optimised C++:

48 kHz    96 kHz    192 kHz

Very high-bandwidth RF processing is where it stops. These need stronger hardware, or FPGA/GPU acceleration:

• tens or hundreds of MHz of complex IQ bandwidth
• full 5G PHY processing
• high-rate beamforming
• large real-time FFT pipelines
• advanced MIMO baseband
• real-time 4K/8K software encoding

Linux is not hard real-time

In practice this means a real-time C++ path should avoid new, delete, malloc and free inside its critical processing loop — allocate once, up front, and reuse.

Power and running cost

Actual power draw

A Pi 5 server with NVMe and active cooling is very efficient:

Annual electricity

At an average of 5 W running continuously:

Cooling

Active cooling is strongly recommended for a 24/7 server. The Raspberry Pi 5 Active Cooler gives lower CPU temperature, less thermal throttling, more stable sustained C++ performance, better Docker behaviour, and a longer comfortable duty cycle. The NVMe SSD also benefits from a small thermal pad or heatsink, especially during long writes.

Software stack

Raspberry Pi OS / Debian
│
├── OpenSSH
├── Samba
├── Tailscale
├── Docker
├── Jellyfin
├── Git
├── rsync
└── optional services

Optional private-cloud applications: Nextcloud, Syncthing, File Browser, Immich, Jellyfin, Gitea, PostgreSQL, Home Assistant.

Security model

Also: SSH keys instead of passwords, root SSH login disabled, security updates installed, unused services firewalled, strong local passwords, and important files backed up.

Backups

RAID is not backup

Primary 4 TB NVMe
       │
       ├── local backup SSD/HDD
       │
       └── encrypted off-site/cloud backup

Important files should exist in at least two independent places.

Backing up the MacBook

rsync -av --progress ~/Documents/ user@raspberrypi:/storage/Backups/MacBook/Documents/

Or mount the SMB share and copy through Finder. A future setup could dedicate a share to Time Machine.

Budget

These are real German retail prices, checked on 24 August 2026 against Geizhals, heise Preisvergleich and billiger.de — not the rough ranges this note originally carried. Availability is good on every line; the figures are the cheapest listed offer, before shipping. The HAT row prices a Pimoroni NVMe Base rather than the GeekPi N04 named in §2.1 — the German comparison sites carry it, it takes the same 2280 drives, and the note's own wording is "or equivalent". Any 2280-capable M-Key HAT lands in the same €15–25 band.

Cutting the cost

Three levers, in order of how much they save and how little they cost you:

• Drop to 2 TB — the Lexar NM790 2 TB is €261.58, saving €187. But 4 TB is the better deal per terabyte (€112/TB against €131/TB), so this only makes sense if 2 TB is genuinely enough for the life of the machine.
• Drop to 4 GB RAM — €137.45, saving €47.45. For a file server, streaming box and Git host, 4 GB is comfortable; 8 GB only starts to matter with several Docker containers or a database under real load (§8.2).
• Use a combined NVMe case — the Argon NEO 5 M.2 NVMe (≈ €36.90) is case and M.2 adapter in one, replacing the €16.50 HAT and the €17.80 case for about €2.60 more, with a tidier result.

A 4 GB + 2 TB build lands at €456 — about a third off — and gives up nothing that matters for a home cloud.

The final configuration

Raspberry Pi 5 — 8 GB
        │
        ├── Official 27 W USB-C PSU
        │
        ├── Active Cooler
        │
        ├── Gigabit Ethernet
        │
        └── GeekPi N04 2280 NVMe HAT
                 │
                 └── 4 TB M.2 2280 NVMe SSD

Raspberry Pi OS 64-bit
OpenSSH · Samba · Tailscale · Docker · Jellyfin · Git · rsync

What it is good at, and what it is not

Upgrade paths

Faster networking

Add a supported 2.5 GbE USB adapter. Practical throughput lands around:

~250–280 MB/s

depending on the adapter, the switch, and system configuration. This is the upgrade that actually moves the bottleneck identified in §4.

More storage

A USB SSD, a USB HDD, a multi-drive enclosure, or a second NAS.

Backup

A second external drive that is disconnected or kept in another place.

Stronger compute

If the workload grows to heavy 5G processing, AI acceleration, multiple 4K transcodes, or large simulations, the answer is a different machine — a mini-PC, a desktop workstation, an NVIDIA Jetson, or an FPGA/SDR platform — while keeping the Pi as the low-power storage and control server.

The whole architecture

The bottom line. For a low-power personal home cloud this is a strong setup: Pi 5 8 GB + active cooling + official 27 W PSU + GeekPi N04 2280 NVMe HAT + 4 TB NVMe SSD + Gigabit Ethernet + Tailscale. It gives roughly 4 TB of local high-speed storage, secure remote SSH, Mac network storage, remote file access, video streaming, Linux/C++ development, Docker, moderate real-time and DSP capability, and very low 24/7 electricity consumption.

For direct high-speed access to the same NVMe from the MacBook, a separate USB4 / Thunderbolt 40 Gbps enclosure can be used — but for normal home-cloud use, keep the SSD permanently installed in the Pi and reach it over the network (§3.2).