diff --git a/halo2_proofs/Cargo.toml b/halo2_proofs/Cargo.toml
index a21b41d9..6a520de9 100644
--- a/halo2_proofs/Cargo.toml
+++ b/halo2_proofs/Cargo.toml
@@ -67,6 +67,9 @@ log = "0.4.17"
 # timer
 ark-std = { version = "0.3.0" }
 
+# binding
+cxx = "1.0"
+
 [dev-dependencies]
 assert_matches = "1.5"
 criterion = "0.3"
@@ -74,6 +77,9 @@ gumdrop = "0.8"
 proptest = "1"
 rand_core = { version = "0.6", default-features = false, features = ["getrandom"] }
 
+[build-dependencies]
+cxx-build = "1.0"
+
 [target.'cfg(all(target_arch = "wasm32", target_os = "unknown"))'.dev-dependencies]
 getrandom = { version = "0.2", features = ["js"] }
 
@@ -87,6 +93,7 @@ shplonk = []
 gwc = []
 phase-check = []
 profile = ["ark-std/print-trace"]
+tachyon_msm_gpu = []
 
 [lib]
 bench = false
diff --git a/halo2_proofs/build.rs b/halo2_proofs/build.rs
new file mode 100644
index 00000000..7da8005d
--- /dev/null
+++ b/halo2_proofs/build.rs
@@ -0,0 +1,23 @@
+fn main() {
+  cxx_build::bridge("src/lib.rs")
+      .files([
+          "src/msm.cc",
+          #[cfg(feature = "tachyon_msm_gpu")]
+          "src/msm_gpu.cc",
+      ])
+      .flag_if_supported("-std=c++17")
+      .compile("halo2_proofs");
+
+  let dep_files = vec![
+      "src/lib.rs",
+      "src/msm.cc",
+      #[cfg(feature = "tachyon_msm_gpu")]
+      "src/msm_gpu.cc",
+      "include/msm.h",
+  ];
+  for file in dep_files {
+      println!("cargo:rerun-if-changed={file}");
+  }
+
+  println!("cargo:rustc-link-lib=dylib=tachyon");
+}
diff --git a/halo2_proofs/include/msm.h b/halo2_proofs/include/msm.h
new file mode 100644
index 00000000..100c8df5
--- /dev/null
+++ b/halo2_proofs/include/msm.h
@@ -0,0 +1,26 @@
+#ifndef HALO2_PROOFS_INCLUDE_MSM_H_
+#define HALO2_PROOFS_INCLUDE_MSM_H_
+
+#include "rust/cxx.h"
+
+namespace tachyon {
+namespace halo2 {
+
+struct CppG1Affine;
+struct CppG1Jacobian;
+struct CppFr;
+
+rust::Box<CppG1Jacobian> msm(rust::Slice<const CppG1Affine> bases,
+                             rust::Slice<const CppFr> scalars);
+
+void init_msm_gpu(uint8_t degree);
+
+void release_msm_gpu();
+
+rust::Box<CppG1Jacobian> msm_gpu(rust::Slice<const CppG1Affine> bases,
+                                 rust::Slice<const CppFr> scalars);
+
+}  // namespace halo2
+}  // namespace tachyon
+
+#endif  // HALO2_PROOFS_INCLUDE_MSM_H_
diff --git a/halo2_proofs/src/lib.rs b/halo2_proofs/src/lib.rs
index e577a8c0..9a08c65d 100644
--- a/halo2_proofs/src/lib.rs
+++ b/halo2_proofs/src/lib.rs
@@ -19,7 +19,7 @@
 )]
 #![deny(broken_intra_doc_links)]
 #![deny(missing_debug_implementations)]
-#![deny(unsafe_code)]
+#![allow(unsafe_code)]
 // Remove this once we update pasta_curves
 #![allow(unused_imports)]
 #![allow(clippy::derive_partial_eq_without_eq)]
@@ -35,3 +35,50 @@ pub mod transcript;
 pub mod dev;
 mod helpers;
 pub use helpers::SerdeFormat;
+
+#[cxx::bridge(namespace = "tachyon::halo2")]
+mod ffi {
+    // Rust types and signatures exposed to C++.
+    extern "Rust" {
+        type CppG1Affine;
+        type CppG1Jacobian;
+        type CppFq;
+        type CppFr;
+    }
+
+    // C++ types and signatures exposed to Rust.
+    unsafe extern "C++" {
+        include!("halo2_proofs/include/msm.h");
+
+        fn msm(bases: &[CppG1Affine], scalars: &[CppFr]) -> Box<CppG1Jacobian>;
+        #[cfg(feature = "tachyon_msm_gpu")]
+        fn init_msm_gpu(degree: u8);
+        #[cfg(feature = "tachyon_msm_gpu")]
+        fn release_msm_gpu();
+        #[cfg(feature = "tachyon_msm_gpu")]
+        fn msm_gpu(bases: &[CppG1Affine], scalars: &[CppFr]) -> Box<CppG1Jacobian>;
+    }
+}
+
+#[repr(C)]
+#[derive(Debug)]
+pub struct CppG1Affine {
+    pub x: CppFq,
+    pub y: CppFq,
+}
+
+#[repr(C)]
+#[derive(Debug)]
+pub struct CppG1Jacobian {
+    pub x: CppFq,
+    pub y: CppFq,
+    pub z: CppFq,
+}
+
+#[repr(transparent)]
+#[derive(Debug)]
+pub struct CppFq(pub [u64; 4]);
+
+#[repr(transparent)]
+#[derive(Debug)]
+pub struct CppFr(pub [u64; 4]);
diff --git a/halo2_proofs/src/msm.cc b/halo2_proofs/src/msm.cc
new file mode 100644
index 00000000..b7b26cc5
--- /dev/null
+++ b/halo2_proofs/src/msm.cc
@@ -0,0 +1,20 @@
+#include "halo2_proofs/include/msm.h"
+
+#include <tachyon/c/math/msm/msm.h>
+
+#include "halo2_proofs/src/lib.rs.h"
+
+namespace tachyon {
+namespace halo2 {
+
+rust::Box<CppG1Jacobian> msm(rust::Slice<const CppG1Affine> bases,
+                             rust::Slice<const CppFr> scalars) {
+  auto ret = tachyon_msm_g1_point2(
+      reinterpret_cast<const tachyon_bn254_point2*>(bases.data()),
+      bases.length(), reinterpret_cast<const tachyon_bn254_fr*>(scalars.data()),
+      scalars.length());
+  return rust::Box<CppG1Jacobian>::from_raw(reinterpret_cast<CppG1Jacobian*>(ret));
+}
+
+}  // namespace halo2
+}  // namespace tachyon
diff --git a/halo2_proofs/src/msm_gpu.cc b/halo2_proofs/src/msm_gpu.cc
new file mode 100644
index 00000000..3841b44a
--- /dev/null
+++ b/halo2_proofs/src/msm_gpu.cc
@@ -0,0 +1,26 @@
+// clang-format off
+#include <tachyon/c/math/msm/msm_gpu.h>
+// clang-format on
+
+#include "halo2_proofs/src/lib.rs.h"
+#include "halo2_proofs/include/msm.h"
+
+namespace tachyon {
+
+namespace halo2 {
+
+void init_msm_gpu(uint8_t degree) { tachyon_init_msm_gpu(degree); }
+
+void release_msm_gpu() { tachyon_release_msm_gpu(); }
+
+rust::Box<CppG1Jacobian> msm_gpu(rust::Slice<const CppG1Affine> bases,
+                                 rust::Slice<const CppFr> scalars) {
+  auto ret = tachyon_msm_g1_point2_gpu(
+      reinterpret_cast<const tachyon_bn254_point2*>(bases.data()),
+      bases.length(), reinterpret_cast<const tachyon_bn254_fr*>(scalars.data()),
+      scalars.length());
+  return rust::Box<CppG1Jacobian>::from_raw(reinterpret_cast<CppG1Jacobian*>(ret));
+}
+
+}  // namespace halo2
+}  // namespace tachyon
diff --git a/halo2_proofs/src/poly/kzg/commitment.rs b/halo2_proofs/src/poly/kzg/commitment.rs
index aa86fdc1..f48e889d 100644
--- a/halo2_proofs/src/poly/kzg/commitment.rs
+++ b/halo2_proofs/src/poly/kzg/commitment.rs
@@ -295,11 +295,24 @@ where
         poly: &Polynomial<E::Scalar, LagrangeCoeff>,
         _: Blind<E::Scalar>,
     ) -> E::G1 {
-        let mut scalars = Vec::with_capacity(poly.len());
+        let mut scalars: Vec<E::Scalar> = Vec::with_capacity(poly.len());
         scalars.extend(poly.iter());
         let bases = &self.g_lagrange;
         let size = scalars.len();
         assert!(bases.len() >= size);
+        #[cfg(feature = "tachyon_msm_gpu")]
+        unsafe {
+            use crate::{ffi, CppFr, CppG1Affine};
+            use std::mem;
+
+            let bases: &[CppG1Affine] = mem::transmute(bases.as_slice());
+            let scalars: &[CppFr] = mem::transmute(scalars.as_slice());
+
+            let ret = ffi::msm_gpu(bases, scalars);
+            let ret: Box<E::G1> = mem::transmute(ret);
+            *ret
+        }
+        #[cfg(not(feature = "tachyon_msm_gpu"))]
         best_multiexp(&scalars, &bases[0..size])
     }
 
@@ -337,11 +350,24 @@ where
     }
 
     fn commit(&self, poly: &Polynomial<E::Scalar, Coeff>, _: Blind<E::Scalar>) -> E::G1 {
-        let mut scalars = Vec::with_capacity(poly.len());
+        let mut scalars: Vec<E::Scalar> = Vec::with_capacity(poly.len());
         scalars.extend(poly.iter());
         let bases = &self.g;
         let size = scalars.len();
         assert!(bases.len() >= size);
+        #[cfg(feature = "tachyon_msm_gpu")]
+        unsafe {
+            use crate::{ffi, CppFr, CppG1Affine};
+            use std::mem;
+
+            let bases: &[CppG1Affine] = mem::transmute(bases.as_slice());
+            let scalars: &[CppFr] = mem::transmute(scalars.as_slice());
+
+            let ret = ffi::msm_gpu(bases, scalars);
+            let ret: Box<E::G1> = mem::transmute(ret);
+            *ret
+        }
+        #[cfg(not(feature = "tachyon_msm_gpu"))]
         best_multiexp(&scalars, &bases[0..size])
     }
 
diff --git a/halo2_proofs/src/poly/kzg/msm.rs b/halo2_proofs/src/poly/kzg/msm.rs
index 6cc90a51..3d13b081 100644
--- a/halo2_proofs/src/poly/kzg/msm.rs
+++ b/halo2_proofs/src/poly/kzg/msm.rs
@@ -66,6 +66,19 @@ impl<E: Engine + Debug> MSM<E::G1Affine> for MSMKZG<E> {
         use group::prime::PrimeCurveAffine;
         let mut bases = vec![E::G1Affine::identity(); self.scalars.len()];
         E::G1::batch_normalize(&self.bases, &mut bases);
+        #[cfg(feature = "tachyon_msm_gpu")]
+        unsafe {
+            use crate::{ffi, CppFr, CppG1Affine};
+            use std::mem;
+
+            let bases: &[CppG1Affine] = mem::transmute(bases.as_slice());
+            let scalars: &[CppFr] = mem::transmute(self.scalars.as_slice());
+
+            let ret = ffi::msm_gpu(bases, scalars);
+            let ret: Box<E::G1> = mem::transmute(ret);
+            *ret
+        }
+        #[cfg(not(feature = "tachyon_msm_gpu"))]
         best_multiexp(&self.scalars, &bases)
     }