Description
In our current GKR-based ZK proof system implementation, we represent data using field elements that occupy 32 bits each. However, in many practical applications, the underlying data is binary in nature—each element effectively represents only 1 bit of meaningful information. This creates an opportunity to optimize memory usage.
Proposed Improvement
We should implement a mechanism similar to LLM quantization where we:
- Store data in a compressed format using 1 bit per element.
- Process computations in 32-bit field elements when necessary.
This approach can substantially reduce the memory footprint for intermediate variables without affecting the computational correctness of our ZK proofs.
Benefits
- Memory Efficiency: Significant reduction in storage requirements.
- Potential Speedup: Smaller data size may lead to faster data transfer and processing.
- Scalability: Improved performance for applications dealing with large binary datasets.
Next Steps
- Prototype the conversion logic between compressed and full-width representations.
- Benchmark memory usage and performance improvements.
- Integrate the solution with existing components and perform end-to-end testing.
- Review and iterate based on the findings.
Description
In our current GKR-based ZK proof system implementation, we represent data using field elements that occupy 32 bits each. However, in many practical applications, the underlying data is binary in nature—each element effectively represents only 1 bit of meaningful information. This creates an opportunity to optimize memory usage.
Proposed Improvement
We should implement a mechanism similar to LLM quantization where we:
This approach can substantially reduce the memory footprint for intermediate variables without affecting the computational correctness of our ZK proofs.
Benefits
Next Steps