Skip to content

Memory System

Node1 edited this page Jun 18, 2026 · 3 revisions

Memory System

Purpose

This page explains XAI OS's memory model for CPU-only AI workloads.

Contents

Why Memory Matters

CPU-only AI is often memory-bound. Effective performance depends on placement, sharing, page-fault behavior, and avoiding duplicate model copies.

Raw DRAM or LPDDR bandwidth is the physical ceiling. Effective CPU-only AI bandwidth is the portion that remains usable after scheduler movement, cache pollution, duplicate weights, page faults, interrupts, and unrelated build/test pressure are removed.

Arenas

XAI OS should use explicit arenas for:

  • shared read-only model weights;
  • private KV/cache state;
  • source-code indexes;
  • tokenizer state;
  • network buffers;
  • build output;
  • logs and telemetry.

No swap should be used for AI memory.

Hugepages and Prefaulting

Model arenas should use hugepages or large pages where supported. Hot arenas should be prefaulted before service READY.

Target invariant:

post-warmup page faults on hot AI paths = 0

NUMA and Bandwidth

On Intel Xeon, a latency-sensitive AI Cell should remain inside one NUMA node by default. Model weights should be NUMA-local or deliberately replicated.

On ARM/NVIDIA N1X-compatible systems, memory policy must account for LPDDR or unified memory behavior and cluster topology.

Admission Control

XAI OS should refuse to start or scale an AI Cell when memory bandwidth pressure would violate existing service contracts.

Platform Notes

Intel Desktop:

  • limited memory channels make duplicate model copies expensive;
  • reduce jitter and background memory pressure before claiming bandwidth gains;
  • keep build/test output away from model and KV/cache arenas.

Intel Xeon:

  • NUMA node locality is mandatory by default;
  • model weights should be node-local or intentionally replicated;
  • memory-channel utilization is a benchmark gate.

ARM/NVIDIA N1X-compatible:

  • LPDDR or unified-memory behavior must be measured directly;
  • cluster locality and memory-controller firmware behavior are risks;
  • SMMU policy must protect DMA into model and KV/cache arenas.

Acceptance Criteria

  • Shared model weights are physically shared and mapped read-only.
  • KV/cache arenas are private per embedded app agent.
  • No swap is used for AI memory.
  • Post-READY hot-path page faults are counted and fail policy unless explicitly allowed.
  • Benchmark output separates raw bandwidth from effective CPU-only AI bandwidth.

Related Pages

Clone this wiki locally