CryptICE™ AI Weight Compression & Encryption IP

Deliver More Model Weights at a Given DDR Bandwidth

Between the DDR controller and on-chip GPU or NPU, CryptICE™ coordinates weight compression, decompression, encryption and decryption. Compression reduces weight-transfer volume, while encryption limits plaintext exposure in off-chip storage and transfer. The IP is ready for volume-production delivery.

Runtime loading compresses and encrypts weights through CryptICE; offline-preprocessed weights enter the write path directly. Both pass through the on-chip DDR controller to off-chip DDR, then return through CryptICE decryption and decompression to supply GPU or NPU inference.

Flexible Weight Loading

Runtime Loading

Compress and encrypt original model weights at runtime before writing them to DDR, supporting loading flows that process weights during system operation.

Offline Preprocessing

The toolchain compresses and encrypts weights in advance. The system loads the preprocessed weights directly, reducing runtime preparation.

Both approaches share the DDR-read, decryption and decompression path that supplies weights to GPU or NPU inference.

Effective Weight Bandwidth

Under specific conditions, effective weight bandwidth can improve by approximately 30%. At the same process node and physical bandwidth, this measures the equivalent uncompressed weight data delivered to compute units per unit of time, relative to a design without weight compression.

Actual gains depend on the model, weight precision, access pattern and configuration. Faster weight delivery can reduce compute stalls in bandwidth-limited workloads.

Integration & Adaptation

Data-path integration, performance evaluation and technical support adapt weight compression and encryption to different models, loading flows and chip platforms.

Evaluate Your Model-Weight Data Path

Share your model, weight precision, access patterns, DDR bandwidth and loading approach. We will arrange a technical discussion.

Contact Us

Back to Cryptographic IP Overview