This release is the third major step toward the 1.0.0 release of Zama’s low-level crypto library, Concrete-core. Check out previous blog posts on the topic (V1.0.0-alpha and V1.0.0-beta), where we explain how Concrete-core is designed to experiment and integrate new FHE-related hardware-acceleration with ease!

Today, it is our pleasure to deliver Concrete-core V1.0.0-gamma, with Cuda acceleration! Visit the Github release note for the full list of changes. Overall, this version introduces:

• A new Cuda backend, dedicated to the acceleration of the programmable bootstrap and the keyswitch on Nvidia GPUs (Nvidia GPUs are widely available on compute environments and highly popular in machine learning, which is why we are targeting this brand specifically).
• A split of the Core backend in order to enable multi-platform support.
• Support for serialization: entities now have a serialization version attached to them, so that it’s possible to know which version of Concrete-core generated them. Serialization engines are also implemented.

This release comes with two new APIs on top of Concrete-core: a C API and a WASM API. Both only wrap a subset of Concrete-core at the moment. The C API covers the use case of the Concrete Framework’s Compiler, while the WASM API provides all necessary engines for a client application. We plan to make their generation automatic and to offer full coverage of Concrete-core’s features. The aim is to enable the widest possible range of applications to be built on top of Concrete-core.

In the sections below, more details are given on the new multi-platform support and the Cuda acceleration introduced in this release.

Enabling multi-platform support

From the start, the V1 design has been aimed at supporting a wide variety of platforms and hardware. Adapting Concrete-core V0.1.10 required an extensive restructuring of the code base; in order to see it through safely, we needed a testing infrastructure. This infrastructure was introduced with the V1.0.0-beta release in April 2022, which set us free to proceed with the main restructuring of Concrete-core: multi-backend support. The former `Core` backend is now split into:

- A default backend that does not depend on specialized hardware, unless specific compile-time configuration is performed:

• It is the main backend for encryption, decryption, and key creation.
• It supports a variety of leveled operations on ciphertexts, but nothing involving the FFT (e.g. bootstrapping, Cmux). It should have an FFT support relying on RustFFT in the future.
• The bootstrap key creation can be accelerated with multithreading, relying on the `rayon` dependency.
• Encryption can be performed using `rdseed` and/or `aesni` on x86_64 platforms that support it. Otherwise, a unix seeder can be used in place of `rdseed` and  CSPRNG software in place of `aesni`. This choice is made at compile time, so it is possible to build the default backend for a chosen target.

- An FFTW backend that implements operations involving polynomial multiplication (bootstrap, Cmux, external product) with FFTW acceleration.

- Additionally, a new Cuda backend is introduced that provides GPU acceleration for the bootstrap and keyswitch.

This new structure is represented in the figure below:

Now let’s dive into more details of the new Cuda backend introduced in this release.

Cuda acceleration of TFHE’s programmable bootstrap

Cuda acceleration is now available in Concrete-core. TFHE’s programmable bootstrap is the bottleneck in terms of performance, which is why this is the first operation we’ve targeted for Cuda acceleration. Since it usually comes together with a keyswitch, we also provide a Cuda accelerated version of the keyswitch. In order to cover different use cases, we offer two different implementations of the bootstrap. In both, the bootstrap operation is accelerated via a single Cuda kernel, a function executed on the GPU that runs a set of instructions onto Cuda threads, themselves part of Cuda blocks:

• The Low Latency Bootstrap (available from the `CudaEngine`), which operates on single input LWE ciphertexts. This uses several Cuda blocks per input ciphertext in the bootstrap kernel. It can also operate on vectors of inputs, but it is aimed at launching restricted numbers of bootstraps simultaneously (from 1 to about 10, depending on the chosen cryptographic parameters).
• The Amortized Bootstrap (available in the `AmortizedCudaEngine`), which is very similar to the ones proposed by NuFHE and CuFHE except that it supports more sets of parameters. It uses one block of threads per input LWE ciphertext, and is aimed at accelerating large numbers of bootstraps (> 10).

Below is a comparison between the nuFHE implementation, the Amortized Bootstrap, and the Low Latency Bootstrap. The parameter set is fixed so as to be supported by nuFHE: 32-bit integers are used along with an LWE dimension of 500, a polynomial size of 1024, one GLWE dimension, two levels of decomposition, and a base logarithm of 10 for the decomposition. nuFHE exposes two PBS implementations: one relying on an NTT (in yellow) and another relying on an FFT (in green). The time it takes to execute one bootstrap when launching various amounts of bootstraps at once is compared considering from one up to 10,000 bootstraps launched at once. 

The amortized bootstrap of Concrete-core is plotted in blue and the Low Latency one in red (the latter can only launch a restricted amount of PBS at once, which is why there are only two points on the curve). This figure above shows that for small amounts of bootstraps launched at once, the Low Latency implementation of Concrete performs best. On the other hand, when launching large amounts of bootstraps at once, the nuFHE implementations and the Amortized bootstrap implementation perform similarly. For intermediate amounts of bootstraps, the nuFHE implementation relying on the FFT performs best,  though it only supports a very limited set of cryptographic parameters. Further, these benchmark results were obtained on an Nvidia Tesla V100-SXM2-16GB GPU. One bootstrap using the same parameters on a CPU requires 27 ms (11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz). 

The benchmarking results shown above only relate to one set of cryptographic parameters, so bear in mind that performance varies depending on the parameters chosen. In order to ease the user’s life, we have introduced cost and noise models for the Cuda accelerated operations. Those are being integrated into Zama’s Optimizer and Compiler, which take care of choosing the best parameters and hardware for the user.

Check out our tutorial to see how to start using the Cuda backend.

Summing Up

With this release, we’re getting very close to the final V1 release. We hope this V1.0.0-gamma version will give you the opportunity to try out your applications with Cuda acceleration. Here are the links to both our user documentation and the Rust documentation to get you started. 

We’re excited to see what you build!

Additional Links

- Release notes

- Github repo

- Documentation

- List of contributors