Links

Run on GPU

TFHE-rs now includes a GPU backend, featuring a CUDA implementation for performing integer arithmetics on encrypted data. In what follows, a simple tutorial is introduced: it shows how to update your existing program to use GPU acceleration, or how to start a new one using GPU.

Prerequisites

  • Cuda version >= 10
  • Compute Capability >= 3.0
  • gcc >= 8.0 - check this page for more details about nvcc/gcc compatible versions
  • cmake >= 3.24
  • Rust version - check this page

Importing to your project

To use the TFHE-rs GPU backend in your project, you first need to add it as a dependency in your Cargo.toml.
If you are using an x86 machine:
tfhe = { version = "0.5.2", features = [ "boolean", "shortint", "integer", "x86_64-unix", "gpu" ] }
If you are using an ARM machine:
tfhe = { version = "0.5.2", features = [ "boolean", "shortint", "integer", "aarch64-unix", "gpu" ] }
When running code that uses TFHE-rs, it is highly recommended to run in release mode with cargo's --release flag to have the best possible performance

Supported platforms

TFHE-rs GPU backend is supported on Linux (x86, aarch64).
OS
x86
aarch64
Linux
x86_64-unix
aarch64-unix*
macOS
Unsupported
Unsupported*
Windows
Unsupported
Unsupported

A first example

Configuring and creating keys.

In comparison with the CPU example, the only difference lies into the key creation, which is detailed here
Here is a full example (combining the client and server parts):
use tfhe::{ConfigBuilder, set_server_key, FheUint8, ClientKey, CompressedServerKey};
use tfhe::prelude::*;
fn main() {
let config = ConfigBuilder::default().build();
let client_key= ClientKey::generate(config);
let compressed_server_key = CompressedServerKey::new(&client_key);
let gpu_key = compressed_server_key.decompress_to_gpu();
let clear_a = 27u8;
let clear_b = 128u8;
let a = FheUint8::encrypt(clear_a, &client_key);
let b = FheUint8::encrypt(clear_b, &client_key);
//Server-side
set_server_key(gpu_key);
let result = a + b;
//Client-side
let decrypted_result: u8 = result.decrypt(&client_key);
let clear_result = clear_a + clear_b;
assert_eq!(decrypted_result, clear_result);
}

Setting the keys

The configuration of the key is different from the CPU. More precisely, if both client and server keys are still generated by the Client (which is assumed to run on a CPU), the server key has then to be decompressed by the Server to be converted into the right format. To do so, the server should run this function: decompressed_to_gpu(). From then on, there is no difference between the CPU and the GPU.

Encrypting data

On the client-side, the method to encrypt the data is exactly the same than the CPU one, i.e.:
let clear_a = 27u8;
let clear_b = 128u8;
let a = FheUint8::encrypt(clear_a, &client_key);
let b = FheUint8::encrypt(clear_b, &client_key);

Computation.

The server must first set its keys up, like in the CPU, with: set_server_key(gpu_key); . Then, homomorphic computations are done with the same code than the one described here.
//Server-side
set_server_key(gpu_key);
let result = a + b;
//Client-side
let decrypted_result: u8 = result.decrypt(&client_key);
let clear_result = clear_a + clear_b;
assert_eq!(decrypted_result, clear_result);

Decryption.

Finally, the client gets the decrypted results by computing:
let decrypted_result: u8 = result.decrypt(&client_key);

Improving performance.

TFHE-rs includes the possibility to leverage the high number of threads given by a GPU. To do so, the configuration should be updated with Rust let config = ConfigBuilder::with_custom_parameters(PARAM_MULTI_BIT_MESSAGE_2_CARRY_2_GROUP_3_KS_PBS, None).build(); The complete example becomes:
use tfhe::{ConfigBuilder, set_server_key, FheUint8, ClientKey, CompressedServerKey};
use tfhe::prelude::*;
use tfhe::shortint::parameters::PARAM_MULTI_BIT_MESSAGE_2_CARRY_2_GROUP_3_KS_PBS;
fn main() {
let config = ConfigBuilder::with_custom_parameters(PARAM_MULTI_BIT_MESSAGE_2_CARRY_2_GROUP_3_KS_PBS, None).build();
let client_key= ClientKey::generate(config);
let compressed_server_key = CompressedServerKey::new(&client_key);
let gpu_key = compressed_server_key.decompress_to_gpu();
let clear_a = 27u8;
let clear_b = 128u8;
let a = FheUint8::encrypt(clear_a, &client_key);
let b = FheUint8::encrypt(clear_b, &client_key);
//Server-side
set_server_key(gpu_key);
let result = a + b;
//Client-side
let decrypted_result: u8 = result.decrypt(&client_key);
let clear_result = clear_a + clear_b;
assert_eq!(decrypted_result, clear_result);
}

List of available operations

The GPU backend includes the following operations:
name
symbol
Enc/Enc
Enc/ Int
Neg
-
✔️
N/A
Add
+
✔️
✔️
Sub
-
✔️
✔️
Mul
*
✔️
✔️
Div
/
✖️
✖️
Rem
%
✖️
✖️
Not
!
✔️
N/A
BitAnd
&
✔️
✔️
BitOr
|
✔️
✔️
BitXor
^
✔️
✔️
Shr
>>
✖️
✔️
Shl
<<
✖️
✔️
Rotate right
rotate_right
✖️
✔️
Rotate left
rotate_left
✖️
✔️
Min
min
✔️
✔️
Max
max
✔️
✔️
Greater than
gt
✔️
✔️
Greater or equal than
ge
✔️
✔️
Lower than
lt
✔️
✔️
Lower or equal than
le
✔️
✔️
Equal
eq
✔️
✔️
Cast (into dest type)
cast_into
✖️
N/A
Cast (from src type)
cast_from
✖️
N/A
Ternary operator
if_then_else
✔️
✖️
All operations follow the same syntax than the one described in here.

Benchmarks

The tables below contain benchmarks for homomorphic operations running on a single V100 from AWS (p3.2xlarge machines), with the default parameters:
Operation \ Size
FheUint8
FheUint16
FheUint32
FheUint64
FheUint128
FheUint256
cuda_add
103.33 ms
129.26 ms
156.83 ms
186.99 ms
320.96 ms
528.15 ms
cuda_bitand
26.11 ms
26.21 ms
26.63 ms
27.24 ms
43.07 ms
65.01 ms
cuda_bitor
26.1 ms
26.21 ms
26.57 ms
27.23 ms
43.05 ms
65.0 ms
cuda_bitxor
26.08 ms
26.21 ms
26.57 ms
27.25 ms
43.06 ms
65.07 ms
cuda_eq
52.82 ms
53.0 ms
79.4 ms
79.58 ms
96.37 ms
145.25 ms
cuda_ge
104.7 ms
130.23 ms
156.19 ms
183.2 ms
213.43 ms
288.76 ms
cuda_gt
104.93 ms
130.2 ms
156.33 ms
183.38 ms
213.47 ms
288.8 ms
cuda_le
105.14 ms
130.47 ms
156.48 ms
183.44 ms
213.33 ms
288.75 ms
cuda_lt
104.73 ms
130.23 ms
156.2 ms
183.14 ms
213.33 ms
288.74 ms
cuda_max
156.7 ms
182.65 ms
210.74 ms
251.78 ms
316.9 ms
442.71 ms
cuda_min
156.85 ms
182.67 ms
210.39 ms
252.02 ms
316.96 ms
442.95 ms
cuda_mul
219.73 ms
302.11 ms
465.91 ms
955.66 ms
2.71 s
9.15 s
cuda_ne
52.72 ms
52.91 ms
79.28 ms
79.59 ms
96.37 ms
145.36 ms
cuda_neg
103.26 ms
129.4 ms
157.19 ms
187.09 ms
321.27 ms
530.11 ms
cuda_sub
103.34 ms
129.42 ms
156.87 ms
187.01 ms
321.04 ms
528.13 ms