Dashboard Overview

ELI5: AMD's term for a single chip that contains both the computer's CPU manager and GPU math speed-runner. Strix Halo shares the system's 128 GB memory pool between CPU and GPU.

Analogy: A premium kitchen machine that blends and cooks on the same counter instead of buying two separate appliances.

ELI5: A system where the CPU and GPU share the exact same physical system RAM pool.

Analogy: A chef and assistant sharing a single massive counter space instead of running between separate tables.

ELI5: The operating system setting that controls how much shared RAM the GPU is allowed to access for graphics and compute allocations. This guide's reference setup uses a 96 GB GTT pool on a 128 GB machine.

Analogy: A boundary line in a shared room telling the graphics card how much space it can use without taking over the whole house.

ELI5: AMD's GPU compute platform, similar in role to NVIDIA CUDA. It provides the HIP backend used by llama.cpp and other local inference stacks.

Analogy: A bilingual translator bridging the gap between your code and AMD graphics hardware.

ELI5: Vulkan is a graphics/compute API. RADV is Mesa's open-source Vulkan driver for AMD GPUs. In this guide, Vulkan/RADV is the fastest measured default path for local generation rows.

Analogy: Vulkan is the road system; RADV is the road crew keeping the AMD lanes paved.

ELI5: AMD's former open-source Vulkan driver. Prefer Mesa RADV; stale AMDVLK ICD files can quietly make the wrong Vulkan driver load.

Analogy: An old road sign that still sends traffic down the wrong street.

ELI5: A Linux service that applies performance profiles. The accelerator-performance profile can reduce power-management drag during local LLM runs.

Analogy: Telling the plant to use the high-load test profile instead of the energy-saving schedule.

ELI5: A massive autocomplete engine trained on billions of texts to predict the next word.

Analogy: A hyper-smart predictive text keyboard that has read the entire internet.

ELI5: Word-pieces that the AI reads and writes. In English, a token is roughly three-quarters of a word, though exact counts depend on the tokenizer.

Analogy: Cutting text into syllable Lego bricks to build sentences rather than reading individual letters.

ELI5: How fast the model reads your input prompt, measured in tokens per second. Higher is better.

Analogy: How fast a reviewer can read the packet before writing comments.

ELI5: How fast the model writes its response, measured in tokens per second. This is the speed you feel while chatting.

Analogy: How fast the reviewer can dictate the final answer after reading the packet.

ELI5: The size of the AI's active notepad (short-term memory).

Analogy: A notebook where the AI writes down what you said and what it did. If it runs out of pages, it starts forgetting the beginning.

ELI5: Speed-reading math tricks and word-guessing loops to accelerate model processing. On Strix Halo, keep Flash Attention enabled with the equivalent of -fa on / -fa 1.

Analogy: Indexing a book for fast lookups (Attention) and having an assistant draft text for a senior editor to quickly check (Speculative).

ELI5: A self-speculative speed trick where the model has its own built-in head for guessing the next few tokens, so no separate draft model is needed.

Analogy: The senior editor has likely next phrases penciled into the margin and can approve several words at once.

ELI5: A model component trained to guess more than one next token at a time.

Analogy: Instead of predicting one word, it sketches the next short phrase for the main model to check.

ELI5: A formatting script that wraps conversational history (User, Assistant, System) and tool-calling data in structured XML or markdown tags so the AI model understands where thoughts end and tool requests begin.

Analogy: A standardized utility log sheet. No matter who takes the measurements, they write them in the exact same boxes so the state compliance auditor can read them instantly.

ELI5: The file format used by llama.cpp and related tools to store local AI models. A .gguf file contains model weights plus metadata needed for inference.

Analogy: An `.mp3` or `.zip` file specifically optimized for loading AI brains.

ELI5: A compression technique that lowers model decimals to integers to shrink file size.

Common labels: Q4_K_M is a balanced 4-bit quant; Q8_0 is higher-quality 8-bit at roughly 2x the weight size; UD-Q4_K_XL is Unsloth Dynamic 4-bit with higher precision for important layers; BF16 is 16-bit precision and much larger.

Analogy: Saving a huge raw photo as a JPEG. It takes up 70% less space, but looks identical to your eyes.

ELI5: A model design where only a small part of the brain is active for each token. A 30B-A3B model has about 30 billion total parameters but activates about 3 billion per token.

Analogy: A hospital with 8 specialist doctors. For a cold, only the 2 required specialists treat you, keeping it fast and cheap.

ELI5: A model where all parameters are used for every token. A dense 7B model uses all 7 billion parameters every time it writes a token.

Analogy: Every specialist reviews every patient, even routine cases. That can be thorough, but it is slower.

ELI5: GGUF compression formats that keep more detail than 4-bit formats while still fitting local hardware.

Analogy: A larger field notebook with clearer handwriting: more room than the tiny version, but easier to read back accurately.

ELI5: The open-source C++ inference library that powers many local LLM tools. It can run GGUF models through CPU, Vulkan, ROCm/HIP, and other backends.

Analogy: The engine under the hood. Different apps may have different dashboards, but many are driving with this engine.

ELI5: A user-friendly tool for downloading and running local LLMs with commands like ollama run model-name. This repo uses llama-server directly, but Ollama is a common llama.cpp-based path.

Analogy: An appliance wrapper around the engine: easier controls, less manual wiring.

ELI5: A separate model family from Google. In this stack, Gemma 4 31B is a cross-family coding experiment — used for quality verification and second-opinion checks, not as a throughput workhorse. It is a dense model: every token reads all 31B parameters, so decode runs at ~8 tok/s on Strix Halo, much slower than the faster MoE lanes (~46–81 tok/s depending on model and MTP opt-in).

Analogy: A specialist reviewer from a different firm — slower to consult, but valuable for a second opinion on tricky plans. Not someone you route every job to.

ELI5: A large open-weight model (OpenAI). In this stack it is the AMERICAN-ONLY quality/speed lane — the US-origin pick for agencies that may require domestic-only model provenance (the general QUALITY champion, StepFun, is non-US in origin).

Analogy: The trusted domestic-supplier option you keep on hand for customers whose rules say "buy American," even when another vendor tops the leaderboard.

ELI5: A chatbot given tools and a goal, running in a loop until it's finished.

Analogy: Giving an assistant a mouse and keyboard, saying: "Clean this file and let me know when done," instead of just asking for advice.

ELI5: The moment the AI decides to run an external program (e.g., a file reader or bash command) instead of guessing.

Analogy: A chef looking up a recipe in an index rather than trying to remember the measurements.

ELI5: A verification test to prove the agent is executing tools rather than hallucinating answers.

Analogy: Putting a secret word inside a box and asking the agent to open the box and tell you the word. If they echo it back, they successfully used the key.

ELI5: A secure, isolated virtual room inside your computer where the AI agent is allowed to write and run code without risk of altering or damaging your actual operating system.

Analogy: A safety hood in a laboratory. You run chemical reactions inside the hood to contain fumes and spills, protecting the rest of the building.

ELI5: A workflow where a main "Manager" AI agent takes a complex user goal, breaks it into smaller sub-tasks, delegates them to specialized "Worker" sub-agents, and compiles their final outputs.

Analogy: A chief operator coordinating plant maintenance, laboratory testing, and electrical crews rather than trying to do every job himself.

ELI5: A blind A/B comparison where two model answers are shuffled and judged prompt by prompt.

Analogy: A taste test with the labels covered. It helps when normal scores are too close to settle the choice.

ELI5: A reliability rule requiring three clean end-to-end passes instead of one lucky success.

Analogy: Starting equipment three times cleanly tells you more than seeing it start once.

ELI5: The industrial computer network that reads sensor data (flow, pressure, tank levels) and operates mechanical hardware (pumps, valves, chemical feeds) in real-time.

Analogy: The dashboard gauges, gas pedal, and steering wheel of a massive commercial truck.

ELI5: The legal safety limit set by the EPA on the concentration of a chemical or contaminant allowed in public drinking water systems.

Analogy: The legal speed limit on a residential street. Going above it triggers a violation and requires immediate corrective action.

ELI5: A walled-off environment that looks like a full computer to the program inside, but cannot reach your real system or files.

Analogy: A child's playpen — they can move freely inside it without reaching the stairs.

Plant analog: The SCADA training simulator. Same screens, same alarms — but a wrong setpoint doesn't dose finished water.

ELI5: Telling a sandbox: "give the program inside this one specific folder of my real computer, and nothing else."

Analogy: Handing the apprentice a folder of pages instead of a key to the whole filing cabinet.

Plant analog: The data binder you handed the apprentice — they see exactly the pages you put in it, nothing else.

ELI5: Give each person, program, or role only the access they need for the job — nothing more. Default to "no" unless required.

Analogy: A new hire gets keys to their office and the break room — not the server room — until the job requires it.

Plant analog: Not every operator gets the supervisor PIN. Not every contractor gets the master keyring. Each role's access is sized to the role.

ELI5: A locked, encrypted vault built into your OS where passwords and keys are stored. Accessible only with your login.

Analogy: A safe-deposit box at the bank — they hold it, only your signature opens it.

Plant analog: The locked key cabinet behind the supervisor's desk. Credentials live there, not on the workstation.

ELI5: An attack where untrusted text (a document, a PDF, an email) contains hidden instructions the agent reads and follows as if you'd typed them. The most-studied current agentic attack vector.

Analogy: A villain mails your secretary a letter that says "move all funds to account X — signed, the boss." If the secretary trusts the letter, the boss never needed to be involved.

Plant analog: A complaint letter containing "approve a 50% chlorine feed increase" — if your agent reads incoming mail, that sentence is a tool call to it.

ELI5: The "Are you sure?" the agent must ask before running a destructive command. Whitelist safe tools; require confirmation for the rest.

Analogy: The "Do you really want to send this email?" pop-up — slows things by half a second, catches one mistake a month.

Plant analog: The two-key chemical-feed setpoint override. The witness signature on a backwash change. Load-bearing wall, not "slowing the work down."

ELI5: The command that immediately stops the agent, and the saved record of every action it took. Know where both are before you start a run.

Analogy: The off-switch and the cash register tape. Stops the action; tells you what happened.

Plant analog: The E-stop on rotating equipment, and the alarm history. You don't disable either "to clean up the screen."

ELI5: "Supervisor mode" on a Linux/Mac computer. Anything run with sudo has full system authority to change settings the regular account cannot.

Analogy: The supervisor's master keycard that bypasses normal authorization on every door.

Plant analog: The supervisor PIN that bypasses alarm acknowledgements. Agents almost never need it.