Power On
Our trained model is still data on storage. Pressing the power button does not start an AI. It starts a computer.
Electrical power reaches the motherboard, CPU, RAM, storage devices, network hardware, and accelerator hardware. The machine begins the same basic startup process used by other modern computers.
At this point there is still no prompt, no token processing, no transformer inference, and no model generating output. We are below the AI layer.
The Boot Path
1. UEFI / BIOS
Firmware on the motherboard is the first major software layer involved in startup. On a modern system this is usually UEFI, although the older term BIOS is still commonly used.
Firmware initializes enough hardware to continue booting, discovers important devices, performs startup checks, and identifies a bootable device or boot entry.
Our NVMe drive is especially important now: it contains the operating system we want to start.
2. Bootloader
Control passes from firmware to a bootloader. On many Linux systems this may be GRUB.
The bootloader's job is not to run the AI. It locates and loads the Linux kernel and the initial environment needed to begin Linux startup, then transfers control to the kernel.
3. Linux Kernel
Now Linux itself begins to take control of the machine.
CPU
The kernel establishes processor operation, scheduling, interrupts, and the mechanisms needed to run processes.
Memory
Linux discovers and manages system RAM, establishes virtual-memory mechanisms, and prepares memory for processes.
Storage
Storage drivers and filesystems make persistent data available, including the software and model files stored on NVMe.
Networking
The kernel provides the network stack and device support that later allow requests and responses to travel through the system.
GPU / Accelerator
The operating system discovers the accelerator hardware. Vendor drivers and supporting software will make it usable by higher-level AI software.
The kernel does not understand English prompts or neural-network concepts. Its responsibilities are still fundamentally those of an operating system: processors, memory, devices, files, processes, security, and communication.
4. systemd Takes Over User Space
Once the kernel has initialized enough of the system, it starts the first user-space process. On our AlmaLinux-style
reference machine, that is systemd.
This should look familiar from HCAW. systemd becomes the service manager responsible for bringing the
rest of the operating environment into its intended state.
└── systemd (PID 1)
├── storage/filesystem services
├── networking
├── logging/monitoring
├── security/support services
└── AI-related service(s), when configured
5. Drivers Make the Accelerator Usable
Physically installing a GPU or other accelerator does not by itself make AI software able to use it. The operating system needs the appropriate device driver and supporting runtime interfaces.
The driver provides the controlled connection between ordinary software and the accelerator hardware. Higher-level software can request operations, allocate accelerator memory, transfer data, and submit computational work without directly manipulating every transistor or hardware circuit.
The GPU is now available for computation, but availability is not intelligence. We still have a Linux computer with specialized mathematical hardware waiting for work.
6. Ordinary System Services Start
Networking comes up. Filesystems are mounted. Logging and monitoring may start. Security policies are applied. Supporting services become available.
If this machine has been deployed as an AI server, its configuration can also instruct the operating system to start the software that will host the model.
This is the point where our familiar Linux boot story begins to meet the AI-specific story.
The Boundary
We now have a running Linux computer.
The CPU is executing processes. RAM is active. Storage is mounted. Networking is available. The GPU/accelerator has been discovered and its driver is available.
The trained model is still stored on disk.
Now an AI runtime or inference service starts.
That is a useful boundary for HAAW: everything so far has been the startup of the computer that will run the AI. What happens next begins preparing the trained model for execution.
Where We Are
What We Have — and Have Not — Claimed
- Established mechanism
- A conventional computer startup brings firmware, bootloader, kernel, drivers, user space, and services into operation.
- Established mechanism
- The accelerator requires software support before higher-level programs can use it for computation.
- Important boundary
- A running GPU-capable Linux system is not, by itself, a running AI model.
- Still ahead
- Starting the AI runtime, reading the trained model, allocating memory, loading model weights, and preparing for inference.