QUANTA
For seventy years, computers were organized around the CPU. We believe the next one is organized around the model.
$ boot ▍
MODEL-NATIVE ARCHITECTURE · MNA
The cost of LLM inference is not arithmetic — it is movement. So we stop hauling data to compute, and move compute into the data.
of energy in typical LLM decode
goes to moving data, not math†
of effective-bandwidth headroom
opened by PIM and pooled memory†
principles define the
Model-Native Architecture
† Order-of-magnitude estimates from public research; see the paper for the full argument.
THE MEMORY WALL
Processor compute sprints; memory bandwidth strolls — and the gap compounds. LLM inference runs head-first into this wall.
Growth rates use classic architecture-literature magnitudes (Hennessy & Patterson; Wulf & McKee, 1995); illustrative.
THE ENERGY LEDGER
A 64-bit floating multiply costs about 4 pJ; fetching the same 64 bits from off-chip DRAM — 640 pJ. Compute stopped being the bottleneck long ago. Movement is.
Each lit dot = one multiply's energy (4 pJ). Data: M. Horowitz, ISSCC 2014 (45 nm magnitudes); cross-node is illustrative.
NOETON · EXPLODED VIEW
Weights & KV cache in residence: PIM arrays + pooled memory + wafer scale
Deterministic dataflow does the work; the CPU is demoted to housekeeping
LLM as kernel: agents are processes, tools are syscalls
Device → edge → cloud; requests land on the cheapest tier
Natural-language / multimodal intent, not commands
↓ keep scrolling to explode the five layers · click a legend entry, or see the stack overview →
LLM SYSTEMS
Prefill eats compute, decode eats bandwidth — they never belonged on the same chip.
One request: compute-bound prefill and bandwidth-bound decode flow to different hardware domains
Compute-bound prefill and bandwidth-bound decode run in separate hardware domains; one request spans tiers.
Enter the lab ↗Hot cache resides in PIM arrays, cold cache sinks to pooled memory — the longer the context, the more tiering pays.
Enter the lab ↗Requests land on the cheapest tier that satisfies them. The device holds the floor; the cloud serves full capability.
Enter the lab ↗Every trade-off starts from the roofline and the energy ledger — physics, not intuition.
Enter the lab ↗AGENTS
When the LLM becomes the kernel, everything in the OS deserves a rewrite.
Enter the Model-OS series: the OS's next phase change →
Scheduling, memory, context management — a crew of agents created, suspended, and collaborating like processes.
Tools as syscalls: permissions, sandboxing, and audit built into the runtime, not patched on later.
Users express intent against context; the machine resolves it into action — the end of command-style interaction.
TRY IT · L5
No command line, no buttons — click an intent and watch the Model-OS compile it into a plan. (Simulated in-page.)
Click an intent above to begin ▍
FROM HERE TO NOETON
Not science fiction — every step stands on industry facts that already happened.
The Model-OS runs as a runtime on today's hardware: agent scheduling, vector memory, tool gateways. Fellow travelers: MemGPT, AIOS, the agent frameworks.
Inference servers built on HBM-PIM × CXL pooling, with split prefill/decode scheduling. The groundwork exists: Samsung HBM-PIM, SK hynix AiM, the CXL 3.x ecosystem.
The complete machine: tensor dataflow fabric plus intent interface. When decode cost drops another order of magnitude, the interaction paradigm turns the page on its own.
EVIDENCE · NEW IN PAPER V2
Rubin CPX built only for prefill, CloudMatrix 384 wiring memory past the CPU, MCP standardized in a year — we archive the industry's moves against the five principles, one event at a time.