ZeroAI

Purpose-built AI model compression IP, combining software compression with hardware decompression for static model data and hardware compression and decompression for dynamic data, all at extremely low latency.

What it does

AI inference is memory bound. During LLM decode the bottleneck is memory bandwidth, not compute, so the amount of data moved per cycle sets the token rate. ZeroAI removes the waste in that data in real time, freeing both capacity and bandwidth from the memory you already have. It integrates into AI accelerators and memory controllers, for example the digital part of an LPDDR memory controller, and the same technology applies to HBM, GDDR, LPDDR, and DDR.

What it unlocks

more usable memory capacity for higher precision and larger context windows

higher memory bandwidth, increasing tokens per second

lower TCO across memory cost and energy

Static and dynamic data

  • Static model data: software compression with hardware decompression.

  • Dynamic data such as weights, activations, and KV cache: hardware compression and decompression.

  • Lossless by default, so AI model retraining is never required.

The algorithm advantage

The patented ZeroPoint algorithm achieves higher compression than state-of-the-art algorithms while operating at much smaller block sizes (as small as 64 bytes), which is what makes very low, deterministic decompression latency possible on the inference path.

Built on more than ten years of research, with 70 patents issued and 24 pending.

Key specifications

Detailed area, latency, and per-model compression figures are configuration dependent.

Target dataAI model data: weights, activations, KV cache. Static and dynamic.
ArchitectureSoftware compression + hardware decompression for static data; hardware compression and decompression for dynamic data.
Memory supportHBM, GDDR, LPDDR, and DDR.
IntegrationInto AI accelerators and memory controllers, for example the digital part of an LPDDR memory controller.
Block sizeAs small as 64 bytes, enabling low-latency inline operation.
Bandwidth improvementApproximately 20–33% across models and data formats: 17% for INT4, approximately 20% for FP8, 30–33% for BF16/FP16.
Compression ratio vs LZ4 / ZstdINT4: 1.26x vs 1.01 / 1.26
FP4: 1.09x vs 1.01 / 1.09
FP8: 1.26x vs 1.00 / 1.23
BF16/FP16: 1.50x vs 1.00 / 1.29
Decompression throughput / mm²128 GB/s per mm².
Model retrainingNot required.

Roadmap

ZeroAI is part of the AIMX family. The roadmap progresses from a lightweight software-compress / hardware-decompress entry point for edge AI and data centers toward higher performance hardware compression and decompression spanning weights, activations, and KV cache. Each phase expands data type coverage and deployment flexibility. A custom integration IP can be delivered today.

Explore the rest of the portfolio

Streaming

ZeroStream

Link compression for C2C, D2D, DRAM

CXL Links

ZeroConnect

Compressed memory for CXL devices

Storage

ZeroStorage

Standards-based LZ4 acceleration

Overview

All products

Back to the portfolio