MORPH turns AI ASICs into SotA throughput machine for zero-knowledge proof

What is it about?

How far could we push AI ASICs for cryptography? Our previous work CROSS [HPCA'26] first demonstrates the superity of AI ASICs in accelerating Homomorphic Encryption (modular arithmetic on non-prime field). Our new work MORPH [DAC’26] further turns AI ASICs into SotA throughput machine for zero-knowledge proof (modular arithmetic on prime field)! Zero-knowledge Proof relies on high-precision (>256-bit) integer on "prime" field, which invalidates RNS. Directly processing high-precision integer without RNS leads to "quadratic" computational complexity. MORPH [DAC’26] wraps the prime-field computation into an extended "non-prime" field! - quadratic -> linear: enable data to be represented and computed with RNS representation, lowing quadratic algorithmic complexity down to linear! - field conversion as a MatMul: transform the representation conversion between prime field and extended field as a low-precision matrix multiplication for TPU acceleration! Welcome new issues, PR, star, fork - the usage from community is the key power to us! Enjoy MORPH at: Paper: https://lnkd.in/g39U-TV3 Code: https://lnkd.in/gM4fdA-i Tutorial: https://lnkd.in/ePJ7Gg2a

Featured Image

Photo by André François McKenzie on Unsplash

Photo by André François McKenzie on Unsplash

Why is it important?

Verifiable Computing based on Zero-Knowledge Proofs (ZKPs) can help distinguish authentic digital content from AI-generated fakes by cryptographically verifying its provenance. Combined with trusted camera authentication, ZKPs can prove that an image or video originated from a genuine capture rather than an AI-generation process. This could restore trust in digital evidence and enable the reliable use of images and videos in legal proceedings, where authenticity is essential. Large prime-field arithmetic is a latency bottleneck in zero-knowledge proofs, including those used in cryptocurrency systems. Multiplication in a large prime field splits large values into low-precision chunks and multiplies all pairs of chunks, incurring quadratic work and sequential carry propagation. Wrapping multiplication in a prime field within an extended non-prime field allows the use of a Residue Number System (RNS). RNS uses roughly twice the original bit width to express multiplication as independent low-precision operations. However, the result needs to be converted back to the original prime field for modular reduction, which remains costly. MORPH~[\ref{morph}] further adopts BAT and MAT to efficiently maps field-conversion pipeline as low-precision layout-invariant matrix multiplication and parallel vector operations, removing sequential carries.

Perspectives