THE NOETON STACK · L1 / L5

L1 Memory-Centric Substrate

MEMORY-CENTRIC SUBSTRATE

Weights are no longer moved — they live inside compute.

The von Neumann machine treats memory as a passive warehouse: every computation hauls data to the CPU. For large models this is a disaster — weight tensors are huge and nearly static, the KV cache keeps growing, and moving them costs far more energy than computing on them. L1 inverts the hierarchy: weights and the KV cache become the resident center of the machine, with compute units embedded in the memory arrays, working in place.

Three proven technologies form this substrate: processing-in-memory (HBM-PIM, UPMEM) builds arithmetic into the memory banks; CXL pooling decouples capacity from a single chassis so the KV cache can stretch across boxes; wafer-scale integration (Cerebras WSE) shows an entire model can live on one piece of silicon. The Noeton synthesizes them into one unified weight-resident substrate.

HBM-PIM arrays

Arithmetic inside the memory banks; attention runs at zero distance

CXL pooled memory

Capacity decoupled from the box; KV cache stretches elastically

Wafer-scale residence

The whole model on one piece of silicon — no off-chip hauling

Near-memory scheduler

Migrates KV blocks across tiers by access temperature

INTERACTIVE · LIVE

Why must weights live inside memory? Watch the energy race: the left hauls data across a bus for every op; the right computes in place.

01

Weights stay resident

Load once, reside indefinitely — "boot" becomes a session-level concept.

02

Bandwidth is performance

Decode throughput scales directly with substrate bandwidth — verify it in the roofline lab.

03

Capacity decoupled

The context-length ceiling moves from "one card's VRAM" to "pool capacity".

Related lab: Tiered KV cache →