Alex

Some thoughts on FHE performance scaling 3 metrics that matter: - latency (how long it takes to do a single operation) - throughput (how many operations can you do per second) - cost-performance (how many operations can you do per $) For a proper benchmark, you should ensure: - that you bench real world data types: eg 64 bit encrypted integers - support the full range of operations (add, mul, div, comp etc) - that the result matches exactly the plaintext. Error rate should be <2^-128 - security should be at least 128 bits What we have seen on the Zama side so far: - math gave us a 10x improvement. We think we can 5x that and get to 50x. - engineering gave us a 10x improvement - GPU is 10x better than CPU cost-performance wise - FPGAs matches GPUs at 1/10th the cost and clock speed - ASIC is 20x better than FPGA So all in all, we can expect FHE in 2027 to be 1,000,000x better cost-performance wise than in 2020 when we started. What does this mean? Today we are at about 1,000x vs 2020. This allows us to support most EVM use cases today (ETH does 20 tps and Base about 100 tps) 1000x from here means 50,000 tps in FHE for blockchain, enough to run VISA and retail CDBC onchain confidentially. Enough to run all RWAs onchain. Enough to even run a country onchain.

Today marks a major milestone in FHE: the release of the first production-ready, stable version of TFHE-rs. This is what underpins all of our products at @zama Ah, and for those who want to understand how it works, there is a 100 page whitepaper that comes with it, detailing every single algorithm used. Enjoy!