Applications

What sixteen states make possible

The Technology page covered how M16 works. This page is about why it matters. Almost every demanding workload in computing, from AI to film to forecasting the weather, comes down to the same two things: storing enormous amounts of numbers, and moving them fast enough to multiply them. Those are exactly what a stit changes.

M15

M12

M11

M14

M10

16 stits: 16¹⁶ = 18,446,744,073,709,551,616 statesBinary needs 64 bits, one full 64-bit word, to hold the same0111001011110000100101001100000100111011011011100000100001011010

The common thread

One change, felt everywhere

Every application below benefits from the same property. A stit is a single unit that can be in any one of sixteen states, where a bit can only be in one of two. It is not a package of bits. But because one stit can hold one of sixteen values on its own, any number needs far fewer of them, so the same chip stores more and moves less. This improves data transfer rates and bandwidth.

16states per stit

A stit is one physical unit that settles into any one of sixteen states. It is a new kind of unit, not four bits bundled together.

2 vs 8units per byte

Telling apart the 256 possible values of a byte takes eight binary cells, or two stits. The same chip area holds more.

¼the units to move

Moving data costs far more energy than computing on it. Fewer units per value means less to move.

M16 is in development. The figures on this page are stit-for-bit comparisons that follow from the architecture, not benchmarks of finished hardware.

01Artificial intelligence

Bigger models on less hardware

A large language model is mostly a very large pile of numbers, its weights, plus a fast way to multiply them. A 70-billion-parameter model stored at 16 bits per weight is about 140 GB. That is more than any single consumer graphics card holds, so today it gets split across several expensive accelerators or squeezed down until it loses quality.

A binary cell can only hold one of two values. A stit can hold one of sixteen, so the same number needs far fewer of them: a 16-bit weight takes sixteen binary cells, or four stits. Weights that fill about 1.1 trillion binary cells fit in about 280 billion stits. A model that needed a rack can need a server. A model that needed a server can fit in a workstation.

For enterprises

Run frontier-scale models on hardware you own, inside your own walls. Fewer accelerators per model means lower capital cost, and keeping sensitive data on-site stops being a trade against model quality.

For individuals

Models that live in a data center today could run on a desktop or laptop: private, offline, and without a subscription. Your notes, code and photos never have to leave your machine to be useful to an assistant.

Longer context windows

A model remembers a conversation in its KV cache, which grows with every token. A 70B-class model with a 128,000-token context can need roughly 40 GB just for that memory. Denser storage lets the same hardware hold far longer documents, codebases and conversations.

Faster responses

Each new word a model writes means reading every weight from memory again, so generation speed is mostly limited by how fast data moves. Each stit sent is one of sixteen states rather than one of two, so every weight takes a quarter as many transfers to move.

02Data centers

More compute per rack, per watt, per square foot

Data centers are limited less by how much they can calculate than by how much power they can draw and how much heat they can remove. In modern chips, fetching a number from memory can use hundreds of times more energy than the arithmetic done on it. Most of the electricity goes into moving data around.

M16 attacks that cost directly. Fewer units per value means fewer units to charge, read and carry across a wire. That compounds up the whole stack: fewer chips per model, fewer servers per workload, less cooling per server.

Density

Serve the same models from fewer racks, or serve more users from the racks you already have, without building a new facility.

Energy and cooling

Less data movement means less power drawn and less heat to remove. Cooling is a large share of a facility’s bill, and it scales with every watt saved.

Grid and siting

New AI facilities are now limited by how much power the local grid can supply. Doing more work per watt stretches every megawatt already connected.

Cost per answer

When a model fits on fewer accelerators, each response costs less to serve. That makes AI cheaper for everyone downstream.

03Video game graphics

Ray tracing that keeps up, worlds that go further

Ray tracing follows paths of light as they bounce through a scene. Each ray asks the same question millions of times per frame: what does this ray hit? Answering it means searching a tree of the scene’s geometry stored in memory, and the rays scatter in every direction, so the search jumps all over that memory. Ray tracing is held back as much by memory as by math.

Render distance hits the same wall. Everything you can see, every mesh, texture and shadow, has to sit in the graphics card’s memory or stream in fast enough to avoid pop-in. Distant scenery gets cut or blurred because there is no room left for it.

Ray tracing

Denser storage keeps more of the scene’s geometry tree close to the cores, so rays find what they hit with fewer slow trips to memory. That means more rays per pixel, more bounces and cleaner lighting at the same frame rate.

Render distance

When every value needs a quarter as many units, the same memory holds far more of the world at full detail. Mountains, cities and forests can stay sharp to the horizon instead of popping in.

Textures and streaming

Higher-resolution textures fit in the same memory, and assets stream from storage faster when each value takes fewer transfers to move.

Neural rendering

Upscaling, frame generation and ray-tracing denoisers are all small neural networks. They run on matrix multiplication, the exact workload M16 is built for.

04Simulations in labs

Finer grids, faster answers

Computational fluid dynamics (CFD) splits air or water into a grid of tiny cells and solves for pressure, velocity and temperature in each one, step after step. It is how aircraft wings, turbines, engines, heart valves and race cars are designed before anything is built.

Larger meshes

When each pressure and velocity value needs a quarter as many units, a mesh that once needed a cluster can run on a single node, and a cluster can resolve turbulence it previously had to approximate.

Faster solvers

The core of most physics solvers is huge, mostly-empty matrices multiplied again and again. That work is limited by memory speed, which is exactly where moving fewer units pays off.

More design iterations

When each run finishes sooner, engineers test more shapes per day. Better wings, quieter fans and more efficient engines come from more iterations, not just bigger ones.

Beyond fluids

Structural analysis, electromagnetics, molecular dynamics and plasma physics share the same grids and the same matrix math, so the same gains carry over.

05Audio recording and manipulation

More tracks, more effects, less latency

Digital audio is a stream of samples, often 96,000 per second per channel at 24-bit depth. Almost everything done to those samples, including EQ, compression, convolution reverb and amp simulation, is multiply-and-add: the same operation at the heart of AI.

In a recording session, all of it has to happen in the few milliseconds between a musician playing and hearing themselves back. When the computer can’t keep up, engineers raise the buffer size, which adds delay, or freeze tracks so they can no longer be adjusted.

Real-time monitoring

More multiply-and-adds per cycle lets sessions run at smaller buffer sizes, so performers hear themselves with effects and no noticeable delay.

Heavier sessions

Hundreds of tracks, each with a full chain of plugins, without freezing, bouncing or rationing effects to save processing power.

Convolution and spatial audio

Convolution reverb, room modeling and immersive formats multiply every sample by thousands of stored values. Denser, faster math makes these practical everywhere, not only on the biggest machines.

AI audio tools

Stem separation, noise removal and voice restoration are neural networks. On M16 they can run in real time during a session instead of as a slow offline render.

06Video editing

Edit the real footage, not the proxy

A single uncompressed 8K frame with full 10-bit color is about 124 MB. At 24 frames per second, that is roughly 3 GB of data every second, before any effects. Editors work around this with low-resolution proxy files, long renders and waiting on the progress bar.

Video codecs, color grading and effects are math applied to every pixel. Codecs are built on block transforms, color grading is matrix math on every pixel, and modern tools like masking and upscaling are neural networks.

Real-time playback

Scrub through full-resolution, high-bit-depth footage with grades and effects applied, without generating proxies first.

Faster exports

Encoding and rendering are throughput problems. Moving and multiplying more information per cycle shortens the wait between finishing an edit and delivering it.

AI-assisted editing

Object masking, rotoscoping, frame interpolation, upscaling and noise reduction are all neural networks, the workload M16 was designed around.

Longer timelines in memory

More frames cached in the same memory means smoother playback and more room for layered, multi-camera edits.

07Consumer Electronics are Getting Pricey

Keeping Compute Accessible and Capable

The applications we have discussed are single examples that are hit daily by enterprises and consumers alike. However, there is a problem that no matter who you are, it will impact you in some way... price.

Even multi-billion dollar companies are struggling to source hardware to keep up with timelines and budget for compute. Binary hardware that is designed to tackle modern day applications is expensive and some is completely inaccessible to the average consumer. Multistate hardware is designed to handle these complex tasks with less energy, less hardware, and less headache.

Getting There

Gressus is young and small, we are currently prototyping our first multistate hardware but, this is costing us essential developmental runway.

Our Ears are Open

We encourage you to reach out to us about anything. If you are a researcher, engineer, enthusiast, investor, or skeptic, we would love to hear from you at [email protected]

Ways to Help

We want to grow a community of people who are interested in multistate computing and a future where technology benefits everyone. Following us on socials, spreading the word, and even just a quick message to keep it up help us in more ways than you know!

And beyond

Anywhere the math is matrices

If a field runs on large grids of numbers being multiplied together, it is a strong candidate for multi-state compute.

Drug discovery and protein design
New materials and batteries
Robotics and autonomous vehicles
Genomics and medical imaging
Financial risk modeling
Fusion and plasma physics

See it for yourself. The Matrix Lab steps through matrix multiplication on a systolic array and counts the digits binary and M16 each have to process.

Open the Matrix Lab Back to the technology