A hardware root of trust is a security chip in a server or a network device that checks the device's firmware before anything else is allowed to start. If an attacker has changed the firmware, the chip stops it from running. It is called a root of trust because every later check depends on it: the chip is trusted first, and it decides what else to trust.
Why check the firmware? Firmware is the software built into the hardware that starts it and controls it. It runs before the operating system, so the operating system and its security software cannot see a change made there. An attacker who replaces the firmware of a server, or of its baseboard management controller (BMC), can stay hidden even after the disks are wiped and the operating system is installed again. The firmware can also be changed before the device reaches its buyer; the page on supply chain attacks explains how.
Here is how the check works. When a device maker releases firmware, it signs the firmware with a private key that only the maker holds. The root of trust chip stores the matching public key, written into the chip in the factory so that it cannot be changed. At power-on, the chip holds the server's CPU and the BMC in reset, which means they cannot start. It reads the firmware from the flash memory chip where it is stored. It calculates the firmware's hash, a short code calculated from the file's contents, and checks the maker's signature against it. Only if the signature matches does it let the CPU and the BMC start.
Imagine a server in a state government data centre in Lucknow. During a routine restart, the chip reads the server's 32 MB BIOS firmware, the firmware that starts the CPU, and finds that it no longer matches the maker's signature. The chip keeps the server from starting. It restores a clean copy of the firmware from a second, protected flash chip, and checks that copy too. Then it lets the server start, and it sends a record of the event to the security team, who can now investigate how the firmware was changed.
How is this different from a TPM, the trusted platform module found in most servers and laptops? A TPM records what started and can prove it later, but it does not usually stop a bad start. A root of trust checks first and blocks a bad start. It also differs from a secure element, a small locked chip in a bank card or passport that stores keys.
Roots of trust are now built into many servers and data centre switches. A chip company sells them by getting its chip chosen for a server maker's design, a design win. After that, the chip ships in every unit of that model.