From 73d8aebd061618c1a7251cde7f94b924596396ba Mon Sep 17 00:00:00 2001 From: Benoit Chevallier-Mames Date: Tue, 4 Jan 2022 18:32:23 +0100 Subject: [PATCH] docs: add information about precision and exactness closes #1273 --- docs/user/basics/intro.md | 1 + docs/user/explanation/fhe_and_framework_limits.md | 3 +++ 2 files changed, 4 insertions(+) diff --git a/docs/user/basics/intro.md b/docs/user/basics/intro.md index 2df958bcb..c102c02cf 100644 --- a/docs/user/basics/intro.md +++ b/docs/user/basics/intro.md @@ -39,5 +39,6 @@ Concrete is a work in progress, and is currently limited to a certain number of The main _current_ limits are: - **Concrete** only supports unsigned integers - **Concrete** needs the integer to be less than 7 bits (included) +- **Concrete** computations are exact (except a very small probability) for computations on 6 bits or less, and exact at a probability close to 90% for 7 bits computations To overcome the above limitations, Concrete has a [popular quantization](../explanation/quantization.md) method built in the framework that allows map floating point values to integers. We can [use this approach](../howto/use_quantization.md) to run models in FHE. Lastly, we give hints to the user on how to [reduce the precision](../howto/reduce_needed_precision.md) of a model to make it work in Concrete. diff --git a/docs/user/explanation/fhe_and_framework_limits.md b/docs/user/explanation/fhe_and_framework_limits.md index f868f51e5..3e44f9881 100644 --- a/docs/user/explanation/fhe_and_framework_limits.md +++ b/docs/user/explanation/fhe_and_framework_limits.md @@ -28,4 +28,7 @@ As of today, the execution of the FHE program is done locally. Notably, in the c As we explained, we wanted to focus first on cryptographic challenges. Performance has been postponed, and will be tackled in the next release. +### Currently restricted to 7 bits computations + +For the moment, we can only perform computations with 7 bits or less. Furthermore, the exactness of computations is only ensured for 6 bits or less; for 7 bits, the computations are exact with a probability close to 90%. Of course, we are working on increasing this limit, and making the probability of a wrong computation as close to 0% as possible. Don't hesitate to look at our [quantization](quantization.md) section to know how to use smaller integers.