diff --git a/test/external/speed_compare_cuda_nv.py b/test/external/speed_compare_cuda_nv.py
new file mode 100644
index 0000000000..48b7578864
--- /dev/null
+++ b/test/external/speed_compare_cuda_nv.py
@@ -0,0 +1,55 @@
+from tinygrad import Device
+from tinygrad.helpers import getenv, colored
+from extra.optimization.helpers import load_worlds, ast_str_to_lin
+from tinygrad.features.search import bufs_from_lin
+
+# move to helpers?
+def colorize_float(x):
+  ret = f"{x:7.2f}x"
+  if x < 0.75: return colored(ret, 'green')
+  elif x > 1.15: return colored(ret, 'red')
+  else: return colored(ret, 'yellow')
+
+if __name__ == "__main__":
+  ast_strs = load_worlds(filter_reduce=False, filter_novariable=True)
+  cudev = Device["CUDA"]
+  nvdev = Device["NV"]
+
+  # NUM=112 python3 test/external/speed_compare_cuda_nv.py
+
+  single = getenv("NUM", -1)
+  if single != -1: ast_strs = ast_strs[single:single+1]
+
+  average_tm_cuda, average_tm_nv = 0, 0
+  for num,ast in enumerate(ast_strs):
+    # cuda compile
+    culin = ast_str_to_lin(ast, opts=cudev.compiler.compiler_opts)
+    culin.hand_coded_optimizations()
+    cuda_prg = cudev.to_program(culin)
+    cubufs = bufs_from_lin(culin)
+
+    nvlin = ast_str_to_lin(ast, opts=nvdev.compiler.compiler_opts)
+    nvlin.hand_coded_optimizations()
+    nv_prg = nvdev.to_program(nvlin)
+    nvbufs = bufs_from_lin(nvlin)
+
+    # warmup
+    tm_cuda, tm_nv = [], []
+    try:
+      cuda_prg(cubufs, {}, wait=True)
+      for i in range(5): tm_cuda.append(cuda_prg(cubufs, {}, wait=True))
+    except RuntimeError:
+      print("CUDA FAILED")
+      tm_cuda = [1e9]
+
+    try:
+      nv_prg(nvbufs, {}, wait=True)
+      for i in range(5): tm_nv.append(nv_prg(nvbufs, {}, wait=True))
+    except RuntimeError:
+      print("NV FAILED")
+      tm_nv = [1e9]
+    average_tm_cuda += min(tm_cuda)
+    average_tm_nv += min(tm_nv)
+    ratio = min(tm_nv)/min(tm_cuda)
+    print(f"{average_tm_nv/average_tm_cuda:5.2f}x -- {num:4d} {colorize_float(ratio)}  {min(tm_nv)*1e6:7.2f} us", nvlin.name)
+    if ratio > 1.1: print(f"NV slower {ratio}", nvlin.ast, nvlin.applied_opts)
diff --git a/tinygrad/runtime/ops_nv.py b/tinygrad/runtime/ops_nv.py
index 4fa9ae8c12..71894bb1d0 100644
--- a/tinygrad/runtime/ops_nv.py
+++ b/tinygrad/runtime/ops_nv.py
@@ -196,7 +196,7 @@ class NVProgram:
     self.constbuffer_0 = [0] * 88
     self.constbuffer_0[6:12] = [*nvdata64_le(self.device.shared_mem_window), *nvdata64_le(self.device.local_mem_window), *nvdata64_le(0xfffdc0)]
 
-    smem_config = min(shmem_conf * 1024 for shmem_conf in [8, 16, 32, 64, 96] if shmem_conf * 1024 >= self.shmem_usage) // 4096 + 1
+    smem_config = min(shmem_conf * 1024 for shmem_conf in [32, 64, 100] if shmem_conf * 1024 >= self.shmem_usage) // 4096 + 1
     self.qmd = qmd_struct_t(qmd_group_id=0x3f, sm_global_caching_enable=1, invalidate_texture_header_cache=1, invalidate_texture_sampler_cache=1,
                             invalidate_texture_data_cache=1, invalidate_shader_data_cache=1, api_visible_call_limit=1, sampler_index=1,
                             cwd_membar_type=nv_gpu.NVC6C0_QMDV03_00_CWD_MEMBAR_TYPE_L1_SYSMEMBAR, qmd_major_version=3,