diff --git a/Manifest.toml b/Manifest.toml
index 7db629f14f..4da1b90d60 100644
--- a/Manifest.toml
+++ b/Manifest.toml
@@ -68,8 +68,10 @@ uuid = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
 version = "6.0.1"
 
 [[GPUCompiler]]
-deps = ["DataStructures", "InteractiveUtils", "LLVM", "Libdl", "TimerOutputs", "UUIDs"]
-git-tree-sha1 = "1b19d415fc3581ff0ed2f57875fca16b5190060a"
+deps = ["DataStructures", "InteractiveUtils", "LLVM", "Libdl", "Scratch", "Serialization", "TimerOutputs", "UUIDs"]
+git-tree-sha1 = "d36b2fd1cfe01afed8c0126aced9f1646a483dbb"
+repo-rev = "4bb076f"
+repo-url = "https://github.com/JuliaGPU/GPUCompiler.jl.git"
 uuid = "61eb1bfa-7361-4325-ad38-22787b887f55"
 version = "0.7.3"
 
@@ -152,6 +154,12 @@ version = "1.1.0"
 [[SHA]]
 uuid = "ea8e919c-243c-51af-8825-aaa63cd721ce"
 
+[[Scratch]]
+deps = ["Dates"]
+git-tree-sha1 = "92245127815ac0bafbdca65a6e12d7a8c32149ed"
+uuid = "6c6a2e73-6563-6170-7368-637461726353"
+version = "1.0.2"
+
 [[Serialization]]
 uuid = "9e88b42a-f829-5b0c-bbe9-9e923198166b"
 
diff --git a/src/compiler/execution.jl b/src/compiler/execution.jl
index ea466dcc08..fc72f9fffb 100644
--- a/src/compiler/execution.jl
+++ b/src/compiler/execution.jl
@@ -291,24 +291,29 @@ when function changes, or when different types or keyword arguments are provided
 """
 function cufunction(f::F, tt::TT=Tuple{}; name=nothing, kwargs...) where
                    {F<:Core.Function, TT<:Type}
-    ctx = context()
-    env = hash(pointer_from_objref(ctx)) # contexts are unique, but handles might alias
-    # TODO: implement this as a hash function in for CuContext
-
+    dev = device()
+    cache = cufunction_cache[dev]
     source = FunctionSpec(f, tt, true, name)
-    GPUCompiler.cached_compilation(_cufunction, source, env; kwargs...)::HostKernel{f,tt}
+    return GPUCompiler.cached_compilation(cache, cufunction_compile, cufunction_link,
+                                          source; kwargs...)::HostKernel{f,tt}
 end
 
-# actual compilation
-function _cufunction(source::FunctionSpec; kwargs...)
-    # compile to PTX
-    ctx = context()
+const cufunction_cache = PerDevice{Dict{UInt, Any}}((dev)->Dict{UInt, Any}())
+
+# compile to PTX
+function cufunction_compile(@nospecialize(source::FunctionSpec); kwargs...)
     dev = device()
     cap = supported_capability(dev)
     target = PTXCompilerTarget(; cap=supported_capability(dev), kwargs...)
     params = CUDACompilerParams()
     job = CompilerJob(target, source, params)
-    asm, kernel_fn, undefined_fns = GPUCompiler.compile(:asm, job)
+    return GPUCompiler.compile(:asm, job)
+end
+
+# link to device code
+function cufunction_link(@nospecialize(source::FunctionSpec),
+                         (asm, kernel_fn, undefined_fns); kwargs...)
+    ctx = context()
 
     # settings to JIT based on Julia's debug setting
     jit_options = Dict{CUjit_option,Any}()
diff --git a/src/initialization.jl b/src/initialization.jl
index c8991b0b8b..a839b1b8d2 100644
--- a/src/initialization.jl
+++ b/src/initialization.jl
@@ -134,6 +134,7 @@ function __runtime_init__()
         end
     end
 
+    initialize!(cufunction_cache, ndevices())
     resize!(__device_contexts, ndevices())
     fill!(__device_contexts, nothing)