strix-halo-optimizations/docs/architecture.md

Architecture

Script Layers

bin/                    User entry points (thin dispatchers)
  audit                   → scripts/audit/
  monitor                 → scripts/monitor/
  benchmark               → scripts/benchmark/
  optimize                → scripts/optimize/

scripts/                Implementation (sourcing lib/ for shared logic)
  audit/                  System assessment
  monitor/                GPU/system monitoring + metrics logging
  benchmark/              llama-bench via toolbox containers
  optimize/               GRUB, tuned, BIOS guidance, verify, rollback

lib/                    Shared bash libraries (sourced, not executed)
  common.sh               Logging, root checks, confirm prompts, paths
  detect.sh               Hardware/config detection from sysfs + system tools
  format.sh               Color output, human_bytes, status indicators, tables

configs/                Reference configuration templates
data/                   Runtime output (gitignored)
docs/                   Documentation

Scripts source libs as needed: always common.sh first, then detect.sh for hardware detection, then format.sh for formatted output. Not all scripts need all three — rollback.sh only needs common.sh, for example. format.sh depends on color variables defined in common.sh.

Data Flow

/sys/class/drm/card1/device/   ──→  lib/detect.sh  ──→  scripts/audit/
/proc/cpuinfo, /proc/meminfo   ──→      (detect_*)  ──→  scripts/monitor/
/proc/cmdline                   ──→                  ──→  scripts/optimize/
tuned-adm, rpm, lspci          ──→                  ──→  scripts/benchmark/
                                                              │
                                                              ▼
                                                         data/
                                                           ├── audits/*.json
                                                           ├── logs/*.csv
                                                           ├── baselines/*/summary.json
                                                           └── benchmarks/*/summary.json

Unified Memory Model

AMD Strix Halo shares physical RAM between CPU and GPU. Two allocation mechanisms:

Type	Description	Configuration
VRAM (dedicated)	Permanently reserved for GPU framebuffer	BIOS setting (UMA Frame Buffer Size)
GTT (dynamic)	System RAM mapped into GPU address space on demand	Kernel boot params: amdgpu.gttsize, ttm.pages_limit

Optimal for LLM workloads: Minimize VRAM (0.5 GiB), maximize GTT (~60 GiB on 64 GB system). The GPU borrows memory when needed and releases it when idle.

Kernel Parameter Math

The formula (using integer GiB arithmetic):

total_physical_gib = (visible_ram + dedicated_vram) / 1024 / 1024   (integer division)
gtt_gib            = total_physical_gib - 4                          (4 GiB OS reserve)
amdgpu.gttsize     = gtt_gib * 1024                                  (MiB)
ttm.pages_limit    = amdgpu.gttsize * 256                             (4K pages)
iommu              = pt                                                (passthrough)

Actual values depend on how the OS reports memory. On a 64 GB HP ZBook with 32 GB dedicated VRAM:

visible_ram ≈ 31.1 GiB, vram = 32 GiB → total ≈ 63 GiB (integer: 59 GiB GTT)
amdgpu.gttsize = 60416, ttm.pages_limit = 15466496

Run make audit to see the exact computed values for your system. The toolkit computes these dynamically via recommended_gttsize_mib() and recommended_pages_limit() in lib/detect.sh.

Sysfs Paths

Card number is auto-detected by find_gpu_card() in lib/detect.sh (matches AMD vendor ID 0x1002, falls back to first card with mem_info_vram_total). Below uses cardN as placeholder:

Path	Content
/sys/class/drm/cardN/device/mem_info_vram_total	Dedicated VRAM in bytes
/sys/class/drm/cardN/device/mem_info_vram_used	VRAM currently in use
/sys/class/drm/cardN/device/mem_info_gtt_total	GTT allocation in bytes
/sys/class/drm/cardN/device/mem_info_gtt_used	GTT currently in use
/sys/class/drm/cardN/device/gpu_busy_percent	GPU utilization 0-100
/sys/class/drm/cardN/device/hwmon/hwmon*/temp1_input	Temperature in millidegrees C
/sys/class/drm/cardN/device/hwmon/hwmon*/power1_average	Power in microwatts

JSON Output Schemas

system-state.json (from audit --json)

{
  "timestamp": "20260325-120000",
  "hardware": { "cpu_model": "...", "cpu_cores": 16, "cpu_threads": 32, "gpu_name": "...", "gpu_device_id": "1586", "system_ram_kb": 32609248 },  // OS-visible RAM (excludes dedicated VRAM)
  "memory": { "vram_total_bytes": 0, "vram_used_bytes": 0, "gtt_total_bytes": 0, "gtt_used_bytes": 0, "recommended_gttsize_mib": 0, "recommended_pages_limit": 0 },
  "kernel": { "version": "...", "cmdline": "...", "param_iommu": "", "param_gttsize": "", "param_pages_limit": "" },
  "firmware": "...", "tuned_profile": "...", "rocm_version": "...",
  "vulkan": { "driver": "...", "version": "..." },
  "sensors": { "gpu_temp_mc": 0, "gpu_power_uw": 0, "gpu_busy_pct": 0 },
  "toolboxes": [], "stacks": { "ollama": "...", "lmstudio": "...", "llamacpp": "...", "opencode": "..." }
}

summary.json (from benchmark runs)

{
  "results": [
    { "file": "model__backend__fa1.log", "model": "...", "size": "...", "backend": "Vulkan", "test": "pp512", "tokens_per_sec": 548.18, "raw": "548.18 +/- 1.59" }
  ]
}

metrics.csv (from monitor --log)

timestamp,gpu_busy_pct,vram_used_mib,gtt_used_mib,gpu_temp_c,gpu_power_w,cpu_pct,ram_used_mib

Sampled every 2 seconds by default. Pure bash implementation (no subshell forks per sample).