CBS-based FHE compiler with custom ISA and transformers-like parallelism (twitter.com)

A new compiler architecture applies a CBS-based approach to fully homomorphic encryption (FHE) by introducing a custom instruction set and a parallelism model reminiscent of transformers’ attention-style dataflow. Rather than treating homomorphic primitives as monolithic operations, the compiler decomposes ML workloads into a small ISA tailored to ciphertext packing, rotation, and SIMD-like batched arithmetic, then schedules and fuses those instructions to maximize parallel use of ciphertext slots and minimize expensive bootstrapping or rotation overheads. For the AI/ML community this matters because it bridges high-level model structure and low-level FHE constraints, making private inference on large models more practical. The transformer-like parallelism maps attention and matrix ops into many independent, pipelined homomorphic kernels that exploit ciphertext batching and reduced multiplicative depth. Key technical implications include improved throughput/latency trade-offs through instruction fusion and slot-aware scheduling, a clearer boundary for compiler/hardware co-design (custom ISA enables optimizations that general-purpose compilers miss), and a path to scale encrypted inference without prohibitive noise growth. Adoption will hinge on concrete benchmarks and how the ISA handles bootstrapping, noise accumulation, and memory/rotation costs, but the design points toward more systematic, scalable FHE compilation for ML.