diff --git a/src/initialization.jl b/src/initialization.jl
index 4bbf898bac..eb976d8699 100644
--- a/src/initialization.jl
+++ b/src/initialization.jl
@@ -134,7 +134,8 @@ function __init__()
     end
 
     # if we're not running under an external profiler, let CUPTI handle NVTX events
-    if !NVTX.isactive()
+    # XXX: JuliaGPU/NVTX.jl#37
+    if !NVTX.isactive() && !Sys.iswindows()
         ENV["NVTX_INJECTION64_PATH"] = CUDA_Runtime.libcupti
         NVTX.activate()
     end
diff --git a/test/core/profile.jl b/test/core/profile.jl
index 551b0762ce..750975fa6e 100644
--- a/test/core/profile.jl
+++ b/test/core/profile.jl
@@ -79,6 +79,9 @@ let
     @test occursin("cuCtxSynchronize", str)
 end
 
+# JuliaGPU/NVTX.jl#37
+if !Sys.iswindows()
+
 # NVTX markers
 let
     str = sprint() do io
@@ -103,6 +106,8 @@ let
     @test occursin("a range", str)
 end
 
+end
+
 end
 end
 
diff --git a/test/runtests.jl b/test/runtests.jl
index 4ac1d24f98..e02a00b0e0 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -2,6 +2,7 @@ using Distributed
 using Dates
 import REPL
 using Printf: @sprintf
+using Base.Filesystem: path_separator
 
 # parse some command-line arguments
 function extract_flag!(args, flag, default=nothing; typ=typeof(default))
@@ -53,12 +54,12 @@ end
 include("setup.jl")     # make sure everything is precompiled
 
 # choose tests
-const tests = ["core/initialization"]    # needs to run first
+const tests = ["core$(path_separator)initialization"]    # needs to run first
 const test_runners = Dict()
 ## GPUArrays testsuite
 for name in keys(TestSuite.tests)
-    push!(tests, "gpuarrays$(Base.Filesystem.path_separator)$name")
-    test_runners["gpuarrays$(Base.Filesystem.path_separator)$name"] = ()->TestSuite.tests[name](CuArray)
+    push!(tests, "gpuarrays$(path_separator)$name")
+    test_runners["gpuarrays$(path_separator)$name"] = ()->TestSuite.tests[name](CuArray)
 end
 ## files in the test folder
 for (rootpath, dirs, files) in walkdir(@__DIR__)
@@ -310,7 +311,9 @@ try
 
                     # tests that muck with the context should not be timed with CUDA events,
                     # since they won't be valid at the end of the test anymore.
-                    time_source = in(test, ["core/initialization", "base/examples", "base/exceptions"]) ? :julia : :cuda
+                    time_source = in(test, ["core$(path_separator)initialization",
+                                            "base$(path_separator)examples",
+                                            "base$(path_separator)exceptions"]) ? :julia : :cuda
 
                     # run the test
                     running_tests[test] = now()