diff --git a/pyproject.toml b/pyproject.toml
index a10488c7a9..b12335f0e0 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -100,7 +100,7 @@ dependencies = [
   'click>=8.0.0',
   'cmake>=3.22',
   'cytoolz>=1.0.1',
-  'dace>=2.0.0a3',
+  'dace==2.3.7',
   'deepdiff>=8.1.0',
   'devtools>=0.6',
   'factory-boy>=3.3.3',
@@ -476,6 +476,9 @@ url = 'https://gridtools.github.io/pypi/'
 # dace = {index = "gridtools"}
 [tool.uv.sources]
 atlas4py = {index = "test.pypi"}
+dace = [
+  {git = "https://github.com/philip-paul-mueller/dace", branch = "phimuell__new-gpu-codegen-dev"}
+]
 
 # -- versioningit --
 [tool.versioningit]
diff --git a/src/gt4py/next/program_processors/runners/dace/workflow/common.py b/src/gt4py/next/program_processors/runners/dace/workflow/common.py
index 6ef363d924..cabcdc2201 100644
--- a/src/gt4py/next/program_processors/runners/dace/workflow/common.py
+++ b/src/gt4py/next/program_processors/runners/dace/workflow/common.py
@@ -109,6 +109,9 @@ def set_dace_config(
     #  This setting allows to throw an exception if any implicit Copy-Map slips thorugh.
     dace.Config.set("compiler.cuda.allow_implicit_memlet_to_map", value=False)
 
+    # Use the new GPU code generator
+    dace.Config.set("compiler.cuda.implementation", value="experimental")
+
     # In some stencils, for example `apply_diffusion_to_w`, the cuda codegen messes
     #  up with the cuda streams, i.e. it allocates N streams but uses N+1. The first
     #  idea was to use just one stream. However, even in that case the generator
diff --git a/uv.lock b/uv.lock
index 7027f7d4f5..f60aeac53b 100644
--- a/uv.lock
+++ b/uv.lock
@@ -1206,8 +1206,8 @@ wheels = [
 
 [[package]]
 name = "dace"
-version = "2.0.0a3"
-source = { registry = "https://pypi.org/simple" }
+version = "2.3.7"
+source = { git = "https://github.com/philip-paul-mueller/dace?branch=phimuell__new-gpu-codegen-dev#896ec9295298d60e956f31834957b863b4fd9a7b" }
 dependencies = [
     { name = "astunparse" },
     { name = "dill" },
@@ -1223,7 +1223,6 @@ dependencies = [
     { name = "sympy" },
     { name = "typing-extensions" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/c3/f8/2401889078017475ce1293af212b76cfdb8a5771ece179851d441ef363f3/dace-2.0.0a3.tar.gz", hash = "sha256:94cbaac4b1f4ef312d24f4151b0905a0e6292a9cfdf9c8b643dabdb2e95b02fa", size = 6005767, upload-time = "2026-05-11T14:49:16.021Z" }
 
 [[package]]
 name = "debugpy"
@@ -1858,7 +1857,7 @@ requires-dist = [
     { name = "cupy-cuda13x", marker = "extra == 'cuda13'", specifier = ">=14.0" },
     { name = "cupy-rocm-7-0", marker = "extra == 'rocm7'", specifier = ">=14.0" },
     { name = "cytoolz", specifier = ">=1.0.1" },
-    { name = "dace", specifier = ">=2.0.0a3" },
+    { name = "dace", git = "https://github.com/philip-paul-mueller/dace?branch=phimuell__new-gpu-codegen-dev" },
     { name = "deepdiff", specifier = ">=8.1.0" },
     { name = "devtools", specifier = ">=0.6" },
     { name = "factory-boy", specifier = ">=3.3.3" },