TFHE-rs v1.1: Fine-Grained GPU Control and More Operators
Blog post from Zama
Zama's TFHE-rs v1.1 introduces significant enhancements for both GPU and CPU backends, focusing on improved performance and versatility in homomorphic computations. On the GPU side, the update aligns cryptographic parameters with the CPU, reducing computational error probabilities, and enhances multi-GPU support, allowing users to specify GPU usage for optimized performance, achieving nearly 500 encrypted 64-bit additions per second on 8×H100 GPUs. For the CPU backend, new operators expand functionality, supporting scalar operations and introducing a dot product operator, improving execution times for various operations. Additionally, TFHE-rs v1.1 features smarter key generation through "chunked" bootstrapping, facilitating operations in memory-constrained environments and enabling key assembly on more powerful servers. These updates aim to enhance the efficiency and scalability of encrypted computations, with future releases expected to continue this trend.