diff --git a/Cargo.lock b/Cargo.lock
index f8e7a0523..d319c82f2 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1384,9 +1384,9 @@ dependencies = [
 
 [[package]]
 name = "cudarc"
-version = "0.13.9"
+version = "0.17.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "486c221362668c63a1636cfa51463b09574433b39029326cff40864b3ba12b6e"
+checksum = "ff0da1a70ec91e66731c1752deb9fda3044f1154fe4ceb5873e3f96ed34cafa3"
 dependencies = [
  "libloading",
 ]
@@ -2565,7 +2565,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "07033963ba89ebaf1584d767badaa2e8fcec21aedea6b8c0346d487d49c28667"
 dependencies = [
  "cfg-if",
- "windows-targets 0.52.6",
+ "windows-targets 0.53.4",
 ]
 
 [[package]]
@@ -2972,7 +2972,7 @@ version = "0.7.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "77e878c846a8abae00dd069496dbe8751b16ac1c3d6bd2a7283a938e8228f90d"
 dependencies = [
- "proc-macro-crate 1.3.1",
+ "proc-macro-crate 3.4.0",
  "proc-macro2",
  "quote",
  "syn 2.0.101",
diff --git a/Cargo.toml b/Cargo.toml
index 43f2aea9a..a0b824a62 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -82,6 +82,7 @@ tracing-subscriber = { version = "0.3", features = ["env-filter"] }
 uint = "0.8"
 
 ceno_gpu = { path = "utils/cuda_hal", package = "cuda_hal" }
+cudarc = { version = "0.17.3", features = ["driver", "cuda-version-from-build-system"] }
 
 [profile.dev]
 lto = "thin"
diff --git a/ceno_zkvm/Cargo.toml b/ceno_zkvm/Cargo.toml
index 07d1394ac..728756197 100644
--- a/ceno_zkvm/Cargo.toml
+++ b/ceno_zkvm/Cargo.toml
@@ -15,7 +15,7 @@ ceno-examples = { path = "../examples-builder" }
 ceno_emul = { path = "../ceno_emul" }
 ceno_gpu = { workspace = true, optional = true }
 ceno_host = { path = "../ceno_host" }
-cudarc = { version = "0.13.0", features = ["driver", "cuda-version-from-build-system"], optional = true }
+cudarc = { workspace = true, optional = true }
 either.workspace = true
 ff_ext.workspace = true
 gkr_iop = { path = "../gkr_iop" }
diff --git a/gkr_iop/Cargo.toml b/gkr_iop/Cargo.toml
index 248e1e2ce..77bbe6716 100644
--- a/gkr_iop/Cargo.toml
+++ b/gkr_iop/Cargo.toml
@@ -12,7 +12,7 @@ version.workspace = true
 [dependencies]
 bincode.workspace = true
 ceno_gpu = { workspace = true, optional = true }
-cudarc = { version = "0.13.0", features = ["driver", "cuda-version-from-build-system"], optional = true }
+cudarc = { workspace = true, optional = true }
 either.workspace = true
 ff_ext.workspace = true
 itertools.workspace = true
diff --git a/gkr_iop/src/gpu/mod.rs b/gkr_iop/src/gpu/mod.rs
index 8c00e6a20..68fef16b5 100644
--- a/gkr_iop/src/gpu/mod.rs
+++ b/gkr_iop/src/gpu/mod.rs
@@ -27,44 +27,32 @@ pub mod gpu_prover {
         common::{
             basefold::utils::convert_ceno_to_gpu_basefold_commitment,
             buffer::BufferImpl,
+            get_ceno_gpu_device_id,
             mle::{
                 build_mle_as_ceno, ordered_sparse32_selector_gpu, rotation_next_base_mle_gpu,
                 rotation_selector_gpu,
             },
         },
     };
-    use cudarc::driver::{CudaDevice, DriverError};
+
     use once_cell::sync::Lazy;
     use std::sync::{Arc, Mutex, MutexGuard};
 
     pub type BB31Base = p3::babybear::BabyBear;
     pub type BB31Ext = ff_ext::BabyBearExt4;
 
-    pub static CUDA_DEVICE: Lazy<Result<Arc<CudaDevice>, DriverError>> =
-        Lazy::new(|| CudaDevice::new(0));
-
     #[allow(clippy::type_complexity)]
     pub static CUDA_HAL: Lazy<
         Result<Arc<Mutex<CudaHalBB31>>, Box<dyn std::error::Error + Send + Sync>>,
     > = Lazy::new(|| {
-        let device = CUDA_DEVICE
-            .as_ref()
-            .map_err(|e| format!("Device init failed: {:?}", e))?;
-        device.bind_to_thread()?;
-
-        CudaHalBB31::new()
+        // can be overridden by env variable `CENO_GPU_DEVICE_ID`
+        let device_id: usize = get_ceno_gpu_device_id(0);
+        CudaHalBB31::new(device_id)
             .map(|hal| Arc::new(Mutex::new(hal)))
             .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>)
     });
 
     pub fn get_cuda_hal() -> Result<MutexGuard<'static, CudaHalBB31>, String> {
-        let device = CUDA_DEVICE
-            .as_ref()
-            .map_err(|e| format!("Device not available: {:?}", e))?;
-        device
-            .bind_to_thread()
-            .map_err(|e| format!("Failed to bind device to thread: {:?}", e))?;
-
         let hal_arc = CUDA_HAL
             .as_ref()
             .map_err(|e| format!("HAL not available: {:?}", e))?;