About Frontier Live core Demo History Future Team Invest Media

Structurally ternary substrate · standard binary CMOS

Ternary chips
you can print today.

A third category of computing — ternary all the way down, printed on the same binary CMOS your foundry already runs. 5× more efficient AI today, on a path to the 12× frontier.

Scroll

The frontier  ·  Efficiency stack

One number is a claim. Three layers is a proof.

There's no single trick. Three independent advantages compound: a structurally ternary substrate, ternary models, and an integrated SoC. Drag the workload from balanced to AI-heavy and watch the envelope move — every layer is graded by evidence.

5.0×
Balanced SKU · measured today

Composite AI performance‑per‑watt versus an ARM Neoverse N3 reference — folded honestly from three measured & published layers.

◀ BalancedAI‑heavy ▶
7.9×fewer gates / MAC vs FP16 · synthesis
−27.3%SoC dynamic power, AI workloads · silicon-grade VCD
5.35×silicon area saving

We fold conservatively — the composite is deliberately smaller than the naïve product of the three layers. Every figure traces to on-disk silicon-grade evidence.

①  Structurally ternary substrateMeasured
1.32× · SoC-aggregate perf/W, per-net VCD · silicon-grade
②  Ternary model architecturePublished
2.0× · BitNet b1.58 ternary weights vs FP16
③  Integrated-SoC envelopeMeasured-proxy
3.7× · neural-render perf/W, measured on GB10 silicon

Live silicon  ·  ternary in motion

Watch a ternary chip think.

Real time, in 3D: weight trits stream into the compute core — +1 passes straight through, −1 flips for free, 0 is skipped entirely — and the result accumulates. Drag to orbit, tap a floating trit, or pulse a batch.

Live compute

Demo  1  ·  Throughput & energy

See the difference.

Same prompt. Same hardware budget. Two architectures. The ratios are what matter — T3ISC delivers the same answer faster while using a fraction of the energy.

Live race
Binary FP16
Conventional baseline (1.0× reference)
LANE_A
Energy use
Live ratio · T3ISC vs FP16
faster throughput
less energy
T3ISC Ternary
Standard CMOS · structurally ternary
LANE_B
Energy use

Ratios derived from measured architectural metrics (energy/MAC, whole-SoC energy, area). Absolute throughput depends on process node, tile size, and workload — see the headline-numbers panel for the underlying figures.

Demo  2  ·  Ternary Data Flow

Same answer. 8× less data.

Each system has a source chamber, a pipe, and a compute chamber. Binary needs the full pipe to deliver every bit. Ternary moves only the trits that matter — same result, 8× less data through the pipe.

Live data flow
Binary today
Full-rate pipe · every bit flows
SOURCE COMPUTE SOURCE COMPUTE
Pipe utilisation
0%
Data through pipe
0 B
Compute fill
0%
T3ISC ternary
8× less data in · same answer out
1/8 SOURCE COMPUTE 1/8 SOURCE COMPUTE
Pipe utilisation
0%
Data through pipe
0 B
Compute fill
0%
less data through the pipe

Why this matters: Ternary software on binary silicon wastes the 8× bandwidth saving. Binary software on ternary silicon blocks the pipeline at the boundary. Only with native ternary silicon + native ternary software do you capture the full advantage end-to-end.

Demo  3  ·  At scale

Same workload. Fewer racks.

Pick a workload. See how many conventional racks it would take versus how many T3ISC racks deliver the same throughput.

Live data centre
Conventional baseline
— racks
vs
T3ISC ternary
— rack
Annual energy saved
Less floor space
Less carbon per inference
Lower running cost

Rack counts are illustrative ratios derived from measured architectural metrics (5.0× whole-SoC energy, 5.35× area). Real deployments depend on process node, memory tier, and workload mix.

About

Ternary chips you can print today.

T3ISC has built ternary computing implemented in standard binary CMOS transistors. Standard fabs. Standard process nodes. Structurally ternary substrate — the architecture is ternary all the way down, but every transistor underneath is the same one your foundry already manufactures by the billions. Printable on any 180 nm-to-3 nm CMOS line today — no specialised process, no exotic devices and no special ternary transistors.

The two existing approaches both fail. Specialised hardware ternary (FeFET, MRAM, multi-Vth) requires non-standard fabs and decade-long device-physics work. Software ternary (BitNet, PrismML) treats ternary as an algorithm only — the model is ternary, the silicon underneath is still binary FP16, and no hardware value is captured.

T3ISC is the third category. Patents filed. Verified on three SIMT platforms (NVIDIA CUDA, Apple Silicon Metal, Linux ARM). 100% simulation-evidence-complete across 16 architectural components:

  • Compute MAC unit
  • TRISC CPU pipeline
  • L1 SRAM (ternary)
  • L1 SRAM (production-mode fallback)
  • L2 cache (3-state)
  • L3 cache (3-state)
  • Weight tier (analog matmul)
  • DRAM (ternary)
  • On-package interconnect
  • Off-package signalling
  • Software stack
  • Lane scheduler
  • Whole-SoC integration
  • TLB
  • Branch predictor
  • Reorder buffer

We are looking for Partners and investors to help us licence this architecture and to bring the future forwards.

Energy per MAC vs FP16
Whole-SoC energy advantage
Silicon area saving
Gates per MAC vs FP16
Less energy per instruction
vs equivalent binary RISC-V on BitNet inference trace

History

From Setun to silicon.

Ternary computing has a 68-year academic history. T3ISC is the first practical commercialisation on standard binary CMOS.

Interactive timeline
1958
Setun
Brusentsov builds the first ternary computer at Moscow State University.
2001
Tristable SRAM
Çilingiroglu publishes the first tristable SRAM cell (IEEE TCAS-II).
2022
IBM HERMES
PCM-based analogue AI inference at ISSCC.
2024
BitNet b1.58
Microsoft proves ternary weights work for transformer LLMs.
2026 · May
T3ISC
Patents filed. Structurally ternary substrate on standard binary CMOS. A new category opens.

The future

What we're building.

From simulation to silicon to scale.

2026 · Foundation

Foundation

  • Patents filed
  • 100% simulation-evidence-complete
  • Full SoC design complete — CPU, memory hierarchy, interconnect
  • Cross-platform validation: NVIDIA CUDA + Apple Silicon Metal + Linux ARM
  • T3ISC ISA software stack · LLVM backend · inference runtime
2027 · Silicon

Silicon

  • Full silicon prototyping
  • Reference SoC bring-up and validation
  • Foundry IP licensing engagements
  • Series A funding round
2028+ · Scale

Scale

  • Datacentre accelerator silicon shipping
  • Mobile and edge AI integration
  • 10× energy-efficiency frontier at production volume
  • Frontier-class LLMs running on-device

Team

Built by engineers.

Wai Chung

Lead Engineer · Inventor

Lead engineer and inventor. Architecture, ISA design, full-stack verification. Inventor and patent filer for the T3ISC substrate.

Join the team

We are hiring senior silicon engineers, ML systems engineers, and full-stack T3ISC backend developers. [email protected]

Engage

Capital and partners.

Investment

  • Series A target: Q4 2026 / Q1 2027
  • Foundation: patents filed, 100% sim-complete, cross-platform validated
  • Use of funds: hardware and foundry partnership engagements, team expansion
[email protected]

Partnership

  • Foundry IP licensing
  • SoC integrator partnerships
  • Strategic integration partners
  • Datacentre AI operators
[email protected]

Get in touch

Email: [email protected]

Media

Press & media.

Press contact

Email: [email protected]

Media kit available on request.

Media enquiry

Stay informed

Get T3ISC announcements first.