CN101 by Normal Computing: status and timeline

CN101

Thermodynamic ASIC for AI inference on PCIe cards. Performs transformer attention in memory, supporting ~1M context windows.

DeployStage 4 of 5

In production and pilot deployment with major semiconductor companies; reference systems available.

Checked 10 Oct0 updates this week

Milestones

Next · In production with major semiconductor customers

CN101 tape-out12 Aug 2025Complete.
CN101 silicon ready for deploymentQ2 2026Complete.
In production with major semiconductor customersTarget Q3 2026Current milestone.
Reference systems deployed and available for pilot programs2026Complete.

Upcoming

Q3 2026In production with major semiconductor customers (next)

Current obstacles

Precision and noise toleranceProving AI workloads maintain accuracy despite noise is essential for deployment.
Competing with GPU ecosystemGPUs dominate; must deliver efficiency, compatibility, and developer familiarity.

Physics limits

Analog circuit yield and variationAnalog circuits are more sensitive to process variation (manufacturing tolerances) than digital logic. Each chip is slightly different, making design and test harder.
Noise is always presentControlling and characterizing noise across temperature and voltage ranges is harder than digital logic.

How it works

3 parts

Noise as a resource

Stochastic analog physics

ASICs use thermal noise to perform computation, trading precision for energy efficiency.

Where data lives

Computation in memory

Attention is computed inside memory, not fetched separately, cutting data movement overhead.

Standard deployment

PCIe integration

Plugs into any server with a PCIe slot; no special infrastructure needed.

Spec sheet

Spec	CN101
Form factor	PCIe accelerator card^R (reported)
Attention context window	~1M tokens^R (reported)
Computation type	Stochastic analog compute-in-memory^R (reported)
Scaling topology	Card to server to rack^R (reported)

^R reported by the company

About Normal Computing

The team behind CN101

Normal Computing

Thermodynamic compute, AI inference accelerators, stochastic analog

Normal Computing builds thermodynamic ASICs for AI inference and AI-accelerated EDA. CN101 is a stochastic analog compute-in-memory accelerator on PCIe cards using thermal noise as a computational resource.

Thermodynamic ASICs use noise as a resource for higher efficiency than conventional logic.
CN101 runs long-context inference on PCIe cards, with attention in memory for ~1M tokens.
Raised $85M+; pilot systems deployed with major semiconductor companies.

normalcomputing.com (opens normalcomputing.com)Company profile →

Founded: 20224 yrs
Headquarters: United States
Status: Private
Valuation: private
Raised: $85M1 round
Last round: Series A · $85M2026
Works in: ComputingAI accelerators & CPUs
Coverage: 1 program · 2 updateslatest 8 Apr 2026checked 10 Oct
Lead investors: Samsung Catalyst

Spec

CN101

Form factor

PCIe accelerator card^R (reported)

Attention context window

~1M tokens^R (reported)

Computation type

Stochastic analog compute-in-memory^R (reported)

Scaling topology

Card to server to rack^R (reported)