Recogni Pareto. The world’s first logarithmic math number system for GenAI

GenAI inference
built to accelerate
the world's AI
ambitions.

Artificial intelligence is math.
Trillions of compute-intensive operations — multiplications and additions.
Every second. Everywhere.

Artificial intelligence is math.
Trillions of compute intensive operations — multiplications and additions. Every second. Everywhere. 

The sheer demand for artificial intelligence computation leads to an explosion in chip size and energy consumption.

The sheer demand for artificial intelligence computation leads to an explosion in chip size and energy consumption.

Introducing
Pareto.

Recogni’s proprietary AI-targeted number system based on the principles of logarithmic math.

By turning multiplications into additions, we reduce computational complexity — making chips a lot smaller, faster, and more efficient.

In standard math, multiplications scale (in power and area) quadratically with the number of bits. Pareto does not use multipliers in hardware — and it consumes far less power and area to begin with. Thus, our ability to deliver high-precision math (e.g. 16 bit) while maintaining low costs is significantly more feasible compared to standard approaches.

Introducing
Pareto.

Recogni’s proprietary AI-targeted number system based on the principles of logarithmic math.

By turning multiplications into additions, we reduce computational complexity — making chips a lot smaller, faster, and more efficient.

In standard math, multiplications scale (in power and area) quadratically with the number of bits. Pareto does not use multipliers in hardware — and it consumes far less power and area to begin with. Thus, our ability to deliver high-precision math (e.g. 16 bit) while maintaining low costs is significantly more feasible compared to standard approaches.

Log numbers are ideal for AI — with the high dynamic range of floating points, and none of the shortcomings.

Log numbers scale with every value.

Error is minimized for small values where it matters most while providing dynamic range for large values.

With virtually no loss of accuracy.

Virtually no loss of accuracy across the board for a range of state-of-the-art AI models. 100.0% is baseline performance — exactly how the model was released in its highest-precision version. Anything below 0.5% is mostly indistinguishable.

Creating a new Pareto-front for accuracy vs. efficiency.

Recogni delivers FP16 accuracy to developers - at a system cost of other FP4 implementations. Pareto literally forms a new Pareto-front in the AI math landscape of accuracy vs. efficiency.

“Bengal cat walking on the moon, smirking into the camera, retro futurism, blue color tone, grainy texture on image, retro nasa”

IEEE FP32 (Baseline)

Recogni Pareto

“Sheep walking on the moon, smirking into the camera, futuristic, blue color tone, grainy texture on image, retro nasa”

IEEE FP32 (Baseline)

Recogni Pareto

“A satellite flying through space, blue planet earth in the background, blue color tone, grainy texture on image, Kodak Tri-X 400, retro nasa”

IEEE FP32 (Baseline)

Recogni Pareto

When applied to image generation, Pareto results show strong consistency with the baseline and maintain high-quality performance.

Log numbers are ideal for AI — with the high dynamic range of floating points, and none of the shortcomings.

Log numbers scale with every value.

Error is minimized for small values where it matters most while providing dynamic range for large values.

With virtually no loss of accuracy.

Virtually no loss of accuracy across the board for a range of state-of-the-art AI models. 100.0% is baseline performance — exactly how the model was released in its highest-precision version. Anything below 0.5% is mostly indistinguishable.

Creating a new Pareto-front for accuracy vs. efficiency.

Recogni delivers FP16 accuracy to developers - at a system cost of other FP4 implementations. Pareto literally forms a new Pareto-front in the AI math landscape of accuracy vs. efficiency.

When applied to image generation, Pareto results show strong consistency with the baseline and maintain high-quality performance.

“Bengal cat walking on the moon, smirking into the camera, retro futurism, blue color tone, grainy texture on image, retro nasa”

IEEE FP32 (Baseline)

Recogni Pareto

“Sheep walking on the moon, smirking into the camera, futuristic, blue color tone, grainy texture on image, retro nasa”

IEEE FP32 (Baseline)

Recogni Pareto

“A satellite flying through space, blue planet earth in the background, blue color tone, grainy texture on image, Kodak Tri-X 400, retro nasa”

IEEE FP32 (Baseline)

Recogni Pareto

And yet, a logarithmic number system has never been implemented in AI.

FP8
FP4
FP8
FP16

No one in the industry has ever achieved an accurate but affordable method of transformation from logarithmic to linear math.


‍
Until now.

And yet, a logarithmic number system has never been implemented in AI.

FP8
FP4
FP8
FP16

No one in the industry has ever achieved an accurate but affordable method of transformation from logarithmic to linear math.


‍
Until now.

Pareto will change how the world runs AI.

Support for the largest GenAI models

Maximum utilization of data center space and energy in the industry

Faster compute than current and upcoming alternatives

Fastest model deployment

Lowest Total Cost of Ownership

No need for liquid cooling

Pareto SDK is now available.

Developer? Join the Waitlist.
Thank you!
Please check your inbox to confirm the double opt in.
Oops! Something went wrong while submitting the form.
For Hyperscalers / For Cloud Service Providers / For Enterprises /

We are building the world’s most efficient exascale multimodal GenAI inference system for data centers