This repo provides demo applications implemented on TFHE-rs, and enables FPGA acceleration on it.
Check out the How to migrate your code for FPGA acceleration? section below to migrate your application. The following steps enable Belfort FPGA acceleration of your THFE-rs code:
// Import the Belfort dependency
use tfhe::integer::fpga::BelfortServerKey;
// Generates the FPGA key from server_key
let mut fpga_key = BelfortServerKey::from(&server_key);
// Connect to the FPGAs
fpga_key.connect();
// Accelerates
BFFA
operations with FPGA
set_server_key(fpga_key.clone());
// The rest of your code stays unchangedAWS accounts do not have access to F2 instances by default. You need to file quota increase request for the Running On-Demand F instances service, which you can search for under Service Quotas > Amazon Elastic Compute Cloud (Amazon EC2). Make sure to combine your request with at least 24 vCPU cores, as f2.6xlarge requires 24 vCPUs. The quota increase may take up to a few days to process.
In your communication to AWS, please pay attention that the F2 access permissions are tied to a region. The FPGA image is available in all the F2 instance regions of today, which are us-east-1, us-west-2, ap-southeast-2 and eu-west-2.
Launch an AWS EC2 F2 instance based on our public Amazon Machine Image (AMI).
- AMI: Belfort FPGA Acceleration AMI on the AWS Marketplace.
- This AMI by Belfort is ready-to-use.
- It's free of charge, but AWS EC2 fees apply
- Instance types:
f2.6xlarge/f2.12xlarge/f2.48xlarge
Pick the instance type depending on how much FPGA acceleration you want;
f2.6xlargefor 1 FPGA (requires access to 24 vCPUs)f2.12xlargefor 2 FPGAs (requires access to 48 vCPUs)f2.48xlargefor 8 FPGAs (requires access to 192 vCPUs)
- SSH into your instance with your credentials
ssh -i <id.pem> ubuntu@<instance_public_dns>-
Clone this repo into your AWS instance
-
Run
prepare_env.shfor cloning TFHE-rs and patching it with the Belfort extensions
cd hello-fpga && ./scripts/prepare_env.shYou can run both CPU and FPGA version of the application and compare the execution time differences;
cargo run --release --package example --bin weighted-sumcargo run --release --package example --bin weighted-sum --features fpgaThe acceleration requires a BelfortServerKey created from the server_key, which connects to the FPGA cores. You can find a weighted-sum example with the code differences for both CPU and FPGA execution below.
Change 5 lines of code:
/// Import dependencies // Import dependencies
| use tfhe::integer::fpga::BelfortServerKey;
/// Create Keys // Create Keys
let config = ConfigBuilder::default().build(); let config = ConfigBuilder::default().build();
let client_key = ClientKey::generate(config); let client_key = ClientKey::generate(config);
let server_key = client_key.generate_server_key(); let server_key = client_key.generate_server_key();
| let mut fpga_key = BelfortServerKey::from(&server_key);
| fpga_key.connect();
set_server_key(server_key); | set_server_key(fpga_key.clone());
// Compute on encrypted data // Compute on encrypted data
// Disconnect from FPGA
| fpga_key.disconnect();Update your Cargo.toml:
- Change the
tfhedependency to use your local fpga-enabledtfhe-rsrepo:
[dependencies]
tfhe = { path = "../../tfhe-rs/tfhe", features = [
"shortint",
"integer",
"experimental-force_fft_algo_dif4",
] }- Add the
fpgafeature to your application'sCargo.toml:
[features]
fpga = ["tfhe/fpga"]
emulate_fpga = ["tfhe/emulate_fpga"]These are the only changes to your code to enable FPGA acceleration.
This repository also contains more comprehensive demo applications. Below you can find the applications and the related commands. They should be run from the root repository and expects and initialized environment.
The Trivium demo contains the transciphering of trivium into FHE. Below you can find its execution command:
# With FPGA acceleration
cargo run --release --package tfhe-trivium --bin demo-shortint --features fpga
# Without FPGA acceleration
cargo run --release --package tfhe-trivium --bin demo-shortintIf you want to specify the number of FPGA cores to use, you can use the alternative connect_to() instead of the connect() function.
This can be useful for development purposes or distributing access of the resources to multiple users.
let mut fpga_key = BelfortServerKey::from(&server_key);
fpga_key.connect_to(vec![0,1,2,3]); // Specifies connection to FPGA cores with indices 0,1,2 and 3
set_server_key(fpga_key);- Additional commands are available to interact with the FPGA's:
fpga-program: programs the fpga's with the Belfort FPGA image released on AWS. This command is only required if the FPGA's were reset.fpga-reset: Resets the fpga images. This is useful if you kill your app withCtrl+Cwhile it interacts with the FPGAs, and the FPGAs are in a bad state.
- In case you run your programs without the fpga's programmed, you will get segmentation faults.
- Lesser used operations are stubbed out with a software implementation. Our team is continuously replacing them with HW optimized versions.
- Enabling the logger gives you runtime warnings if a non-accelerated function is used. Contact us if you would like priority support for a function that emits a warning.
- Current implementations use FFT, but NTT support is under development.
- Development for a specialized cloud environment with optimized performance is ongoing.
Belfort's AMI is free to use only for development, research, prototyping, and experimentation purposes. However, for any commercial use of Belfort's AMI, companies must purchase Belfort’s commercial AMI license.
This software is distributed under the BSD 3-Clause Clear license. Read the license for more details.