Ternara Semiconductor

Minus one. Zero. Plus one.

Four chip designs. From one tile to thirty-two.

Concept render, not a fabricated device

Ternara: intelligence where there is no connection.

We are developing compute cores for compact devices that process sensor data locally. Our goal is to help monitor equipment, agriculture and remote sites of any kind with Ternara AI, without a constant connection to the cloud.

A multiply becomes a choice.

Each of the four designs passes simulation in CI, then goes through automatic place-and-route and layout checks in an open toolchain. The main compute unit of Φ, e and γ is a 16-bit binary floating-point dot product (GF16); their ternary blocks are test circuits.

The lineup

Φ Ternara Φ The smallest Ternara layout: a reset-time numeric anchor and a Lucas-number self-check in one tile. 1x1 tiles Learn more ›

e Ternara e Mid-size test vehicle: a 2x2 mesh of GF16 dot-product tiles plus on-die ternary matrix self-checks. 8x2 tiles Learn more ›

γ Ternara γ The largest Ternara layout: eight spiking-neuron columns, a GF16 dot-product mesh and the ternary matrix self-checks in 32 tiles. 8x4 tiles Learn more ›

C Ternara Corona Format oracle: a lookup ROM of 80 numeric formats with decoders that turn them into FP32 or INT32. 4x4 tiles Learn more ›

τ Ternara τ Concept: a lane that multiplies 8-bit activations by ternary weights, for BitNet-class language models. Concept Learn more ›

Ternara Φ

One tile. A known answer at reset.

Ternara Φ fits in a single tile. After reset, it outputs a known 16-bit value, and a gate-level test confirms it.

Footprint: 1x1 tile
Standard cells: 532
Reset value: 0x47C0
Gate-level tests: 5 of 5 pass¹
Layout checks: LVS, DRC and antenna: clean²
Not measured: power, signed-off timing

Concept render, not a fabricated device · Layout render from the design files. Not a photograph of silicon.

Ternara Φ layout — Concept render, not a fabricated device · Layout render from the design files. Not a photograph of silicon.

Ternara e

A test design for ternary matrices.

Ternara e is a 16-tile test design with 16x16 and 8x8 ternary matrix blocks, where each product is an XOR and a count, not a multiply. In this version the blocks run once after reset on fixed zero operands and cannot be loaded from the pins.

Footprint: 8x2 tiles
Standard cells: 15,297
Ternary matrix blocks: 16x16 and 8x8³
RTL tests: 3 of 3 pass⁴
Layout checks: LVS, DRC and antenna: clean²
Known RTL defects: 3, fixes not in this layout¹²
Static timing: setup violations in slow corners⁴
Not measured: power, signed-off timing

Ternara e layout — Concept render, not a fabricated device · Layout render from the design files. Not a photograph of silicon.

Ternara γ

Eight columns of neurons.

Ternara γ is our largest design: 32 tiles with eight spiking-neuron columns and ternary matrix test blocks. Logic fills about 17 % of the area.

Footprint: 8x4 tiles⁵
Standard cells: 18,176
Neuron columns: 8³
RTL tests: 5 of 5 pass⁴
Layout checks: LVS, DRC and antenna: clean²
Known RTL defects: shared GF16 code, not re-checked¹²
Static timing: setup violations in slow corners⁴
Not measured: power, signed-off timing

Ternara γ layout — Concept render, not a fabricated device · Layout render from the design files. Not a photograph of silicon.

Ternara Corona

Eighty formats. One index.

The Ternara Corona design holds 80 numeric formats in a ROM. In simulation, a 7-bit index selects a format and its decoders return FP32 or INT32. No die exists or is expected for this layout.

Footprint: 4x4 tiles
Formats in ROM: 80
Decoders: 17⁶
Gate-level tests: 51 of 51 pass
Acceptance vectors: 1,920⁷
Status: GDS + precheck, not submitted⁶
Not measured: power, signed-off timing

Ternara Corona layout — Concept render, not a fabricated device · Layout render from the design files. Not a photograph of silicon.

Ternara τ · Concept

Concept: ternary weights, 8-bit inputs.

Ternara τ is a concept lane for BitNet-class language models. Each 8-bit activation times a ternary weight becomes +x, 0 or -x.

Weights: -1, 0, +1¹⁰
Each product: +x, 0 or -x
Related block, in simulation: 27-trit dot product (ternary x ternary), 200 of 200 vectors match⁹
Status: concept (not built)⁸

Layout

Straight from the file.

Every shape inside the tile edge comes from the layout file: cells, metal and vias. We pick the layer colours and add only a flat background and the tile edge.

Layout render from the design files. Not a photograph of silicon.

Placement

Three designs. One public shuttle index.

Ternara Φ, e and γ are listed in the public index of a multi-project shuttle, next to designs from other teams. An entry in an index is not a die. Corona has not been submitted.

Where the Ternara designs sit in the floorplan of a multi-project shuttle, redrawn from its public index (updated 2026-06-13, read 2026-10-10). A floorplan diagram, not a photo of a die. As of 2026-10-10: no die has been returned from fabrication. No Ternara silicon has been measured.

Ternary weights

Subtract. Skip. Add.

-1Subtract

0Skip

+1Add

A ternary weight needs no multiplier. It picks a sign, or skips the term. The logic stays binary. The weights carry three values.

Binary logic, ternary weights. A trit is stored in two bits of binary CMOS.

Chip passport

Every chip gets a passport.

A number without its conditions proves little. A Ternara passport is the record that travels with every result: exactly which files were built, which die was measured, under what conditions and with what uncertainty. Fields that need a real die stay open until there is one.

Ternara ΦPassport draft for the layout listed in the public shuttle index

TRI-Φ-0001 · DRAFT

Artifact identity: layout file SHA-256 6a709afe…0529a6b
Placement: public shuttle index, address 194, 1×1 tile
Process kit: open-source PDK, open toolchain
Gate-level test: 5 of 5 pass in CI
Interface representation: 16-bit reset value 0x47C0, bit-exact in simulation
Device instance: no die yet
Environment: temperature and supply: not measured
Measurement boundary: power: not measured
Uncertainty: needs repeated readings on a die

Open fields are left open on purpose: they can only be filled from a measured die.

The passport format is a proposal; no standards body has adopted it.

Green AI

Green AI, by arithmetic.

Most of the energy in AI goes into moving weights and multiplying them. Ternary weights attack both: three values instead of thousands, and no multiplier at all. Here is the evidence, with sources.

× → ±

No multiplier

With weights limited to −1, 0 and +1, each multiplication becomes subtract, skip or add. Microsoft's BitNet paper describes its matrix multiplication as needing almost no multiplications. Source: arxiv.org ↗

≈10×

Less weight memory than BF16

A ternary weight carries log2(3) ≈ 1.585 bits, against 16 bits for BF16: about 10× less memory for those weights. Practical llama.cpp packings use 1.6875 or 2.0625 bits per weight. Source: github.com ↗

55–82%

Less CPU energy per token, reported by Microsoft

Microsoft reports that its bitnet.cpp software, running ternary BitNet b1.58 test models, used 55.4–70.0% less energy per token than llama.cpp on an Apple M2 Ultra CPU. On an Intel Core i7-13700H laptop CPU, Microsoft measured 71.9–82.2% less energy per token with bitnet.cpp than with llama.cpp (700M and 7B test models). Source: arxiv.org ↗

1.3–2.6 nJ

One DRAM read, versus 0.03 pJ for an 8-bit add

The same table puts a 64-bit DRAM read at 1.3–2.6 nJ, the cost of tens of thousands of 8-bit additions. Moving fewer bits per weight matters more than the arithmetic. Source: gwern.net ↗

945 TWh

Data centres by 2030, according to the IEA

Data centres used about 415 TWh of electricity in 2024, around 1.5% of the world's total, and are set to reach around 945 TWh by 2030, slightly more than Japan uses today. Source: iea.org ↗

As of 2026-10-10: no die has been returned from fabrication. No Ternara silicon has been measured. Power has never been measured on any Ternara hardware. Every energy number on this page belongs to other teams' software, models or chips. None of them describes Ternara. Microsoft's CPU energy figures compare bitnet.cpp with llama.cpp on test ("dummy") models, not trained ones. The 71.4× arithmetic saving and the 0.028 J per token are model estimates, not measurements.

Vision

Smart devices

Smart without the cloud.

Tell an air conditioner “cool the room to 22, quietly” and it understands and does it. No server, no internet. The air conditioner is just an example: a language model on ternary weights could live inside any device. And it matters most where there is no internet at all.

Where there is no internet

Cloud AI works only while there is a connection. In the taiga, in the mountains, at sea, in a remote village there is none, and the smart assistant goes silent. A model that lives inside the device itself needs no tower, no satellite and no server: it understands and answers on the spot.

2.2 billion

people worldwide still do not use the internet, according to ITU figures for 2025. Source: itu.int ↗

58%

of people in rural areas use the internet, against 85% in cities, according to ITU. Source: itu.int ↗

96%

of those offline live in low- and middle-income countries, according to ITU. Source: itu.int ↗

Devices that need no connection

Field translator “Tell the medic: the child has had a fever since yesterday”

Village school tablet “Explain fractions using apples”, a lesson with no internet

Mechanic's manual “The diesel won't start in the frost. What do I check, in order?”

Farm controller “Water the north plot if the soil is dry, leave the greenhouse alone”

Solar water pump “How much water is left in the tank, and will it last till evening?”

Ship's logbook “Log today's catch and how much fuel is left”

Expedition terminal “Merge the notes from today's three sites into one report”

Helper for the elderly “Remind me to take my medicine at eight tonight”

These are examples of where such a model could go, not finished products. Understanding and answering happen inside the device; sending anything out still needs a connection.

How it works inside the device

1. Hears A microphone and a speech recognizer turn the voice into text right inside the device.
2. Understands A small language model on ternary weights works out the meaning: what to do, how much, when, how.
3. Acts The model hands the appliance a plain command, such as “twenty-two degrees, quiet mode”, and the appliance carries it out.
4. Nothing leaves Voice, text and decision stay inside the device. No cloud means no round trip, no server fees and no conversations leaking out.

At home too: from your phone to your washing machine

Smartphone “Sum up my chat with Anna and remind me about the meeting”, offline

Computer “Find the contract where we discussed deadlines and list the key points”

Game console Characters answer anything you say in natural speech, with no server

TV “Play that film where the hero survives alone on Mars”

Washing machine “Wash the wool sweater gently and have it done by seven”

Oven “Heat to two hundred and switch off in forty minutes”

Car “I'm cold, make it warmer and turn on the seat heater”

Robot vacuum “Clean under the kitchen table, stay out of the bedroom”

Headphones “Make the person I'm talking to louder and the café quieter”

Home lighting “Make the light warm and dim it for the night”

Factory machine “Why did the line stop?”, answered in plain words

Children's toy “Tell me a story about space”, and the child's voice never goes online

These are examples of where such a model could go, not a list of finished products.

Why ternary weights

Less memory Every model weight is only −1, 0 or +1. In our blocks such a weight is stored in two bits, where a usual weight takes sixteen. The same memory holds a model eight times larger.
Simpler arithmetic Multiplying by −1, 0 or +1 means subtract, skip or add. No large multiplier is needed, so the block is easier to fit into an appliance's low-cost chip.
Fewer memory reads Fewer bits means less data moved out of memory, which is usually where the energy goes. The power draw of Ternara blocks has not been measured yet.

Understands, then actsFree-form speech instead of a fixed command list.
Works offlineNo outage when the internet is down.
PrivateVoice is processed in the room where it was spoken.
Fits in less memoryVision: ternary weights could fit a larger language model into an appliance's memory, so a device follows free-form speech instead of a fixed command list. Nobody has shown this in an appliance, and no Ternara block is in any device.

Already working today

In 2019 Google shipped an 80 MB speech recognizer that runs on Pixel phones in Gboard and stays available offline. Source: research.google ↗
Google's Gemini Nano runs inside Android's AICore service and can answer without a network connection or sending data to the cloud. Source: developer.android.com ↗
Espressif's MultiNet recognizes up to 200 spoken commands offline on an ESP32-S3 chip, within 500 ms. Its documentation uses "turn on the air conditioner" (in Chinese) as an example command. Source: docs.espressif.com ↗
TCL sells the VoxIN air conditioner with an offline voice module that accepts voice commands without an internet connection. Source: tcl.com ↗
Microsoft reports that bitnet.cpp runs a 100B-parameter BitNet b1.58 test model on a single CPU at 5–7 tokens per second, about human reading speed. Source: github.com ↗

What it would take

Offline voice control is not new: appliances with fixed command lists already ship. A free-form language model inside an appliance has not been shown.

Vision

What's next: networks made of devices.

How devices could pass data to each other without a tower on every block, and earn from it, is on a separate vision page.

See the vision →

Compare the designs.

Φ Ternara Φ e Ternara e γ Ternara γ C Ternara Corona

Footprint 1x1 tile 8x2 tiles 8x4 tiles 4x4 tiles

Standard cells 532 15,297 18,176 1,356

Role Reset-value test design Ternary matrix test design Neuron columns, test design Numeric-format oracle

Simulation in CI 5 of 5 gate-level 3 of 3 RTL 5 of 5 RTL 51 of 51 gate-level

Layout checks LVS, DRC and antenna: clean LVS, DRC and antenna: clean LVS, DRC and antenna: clean LVS and DRC: clean

Public index Listed Listed Listed GDS + precheck, not submitted

Measured on a die No No No No die expected

Not measured: power, signed-off timing

Designed by Ternara.

Specified and simulated by the Ternara team; placed and routed with an open toolchain. Fabrication is a separate, external step. The Technical page names it and shows where each design stands.

Put ternary compute to work.

Ternara is a technology, not just a chip: compute blocks for ternary-weight AI, taken from a written spec to a finished layout, with evidence at every step. Start with a pilot: one block for your chip, device or product, with its checks and a chip passport.

Request a pilot