IPMI is a small computer inside your server that runs whether or not the server itself is running. It has its own network connection, its own power, and its own login, which is what lets you reach a machine that will not boot.
Vendors brand it differently: iDRAC on Dell, iLO on HP, IPMI generically. The capabilities are the same.
What it lets you do
Power control. Power on, power off, and a hard reset. The equivalent of holding the button, without anyone walking to the rack.
Console access. A remote view of the screen as if you were standing at it, including the boot process and the BIOS.
Virtual media. Attach an installation image over the network and boot from it, which is how you reinstall an operating system remotely.
Hardware status. Temperatures, fans, power supplies, memory errors, read directly from the hardware rather than from the operating system.
Why it matters more than it seems
The difference is between problems you can fix and problems that require someone else.
SSH needs the operating system running and the network stack working. When a kernel update leaves the machine unbootable, when a firewall rule locks you out, when the filesystem needs a repair before it will mount, SSH is gone and IPMI is not.
Without it, every one of those is a support ticket and a wait. With it, they are ten minutes.
Test it before you need it
The single most useful thing in this article.
On the day the server is provisioned, log in to IPMI, open the console, and confirm you can see the screen. Confirm you have the password and that it works.
Doing this at three in the morning, on a machine that is already down, is how people discover the password was never sent, the console needs a browser plugin they do not have, or the interface is not reachable from where they are.
Ten minutes on a working day. It is the cheapest insurance a dedicated server has. Provisioning and initial setup goes into what else belongs in that first hour.
It must not face the internet
This is the part that gets people compromised, and it deserves to be stated bluntly.
IPMI implementations have a long history of serious vulnerabilities, including authentication bypasses. The firmware is updated rarely and often never.
And access to IPMI is total: whoever holds it can power the machine off, boot it from their own image, and read every disk. It bypasses every control inside the operating system, because it sits beneath the operating system.
So: never expose it to the internet. Put it on a management network, or reach it through a VPN, or restrict it to specific addresses at the provider's firewall. If your provider offers a VPN or jump host for management access, use it.
Change the default password immediately
Default IPMI credentials are published, and scanners look for them continuously.
Change the password on the first login, delete or disable any default account, and store the credential somewhere that does not depend on the server being reachable: a password manager, not a note on the machine itself.
If you have several servers, use different passwords. One shared IPMI password means one exposure compromises every machine you own.
The console is genuinely a console
Everything typed there goes to the machine as if from an attached keyboard, before any network service is involved.
That means it works in single-user mode, at the boot loader, and in the BIOS. It also means a stuck key or a mistyped command has the same effect as it would at the rack.
Detach the console when you have finished. Some implementations allow only one session, and a forgotten one blocks the next person, including you, tomorrow.
Reinstalling remotely
Attach an installation image as virtual media, set the boot order, and reboot. The installation runs on your screen exactly as it would locally.
Two cautions. It is slow. The image is transferred over the management connection, which is not fast. And it will happily wipe the disks, with the same finality as doing it in person.
Confirm your backups are off the machine before you start. Installing and managing operating systems goes into the install itself.
Hardware alerts from beneath the OS
IPMI reports hardware events the operating system may never see: a failing power supply, memory correcting errors, a fan slowing, temperatures rising.
Configure it to email those alerts. A memory module producing correctable errors is a warning weeks before it produces uncorrectable ones, and that is the difference between a scheduled replacement and an outage.
Send them somewhere that does not depend on this server: a warning about a failing machine, delivered by the failing machine, arrives only when it is not needed. Monitoring server health explains the software side.
When the console is blank
A blank console does not mean the machine is dead.
Check the power status in the interface: if it reports powered on, the machine is running and the display is not being driven, which points at the operating system instead of the hardware.
Try a graceful shutdown first, then a hard reset. The hard reset is the equivalent of pulling the power, so it can leave a filesystem needing repair: use it when the graceful option has failed rather than as the first move. There is more on what to do once you can see the screen again in troubleshooting common issues.