diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index 9ccf1d1d74..8213b1cc9a 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -658,7 +658,7 @@ jobs:
           AMD: 0
         run: |
           PYTHONPATH=. NULL=1 EMULATE=AMD python extra/mmapeak/mmapeak.py
-          PYTHONPATH=. NULL=1 EMULATE=AMD_CDNA4 python3 -m pytest -n=auto test/testextra/test_tk.py test/testextra/test_asm_gemm.py
+          PYTHONPATH=. NULL=1 EMULATE=AMD_CDNA4 python3 -m pytest -n=auto test/testextra/test_tk.py test/backend/test_asm_gemm.py
       - name: Run ASM matmul on MOCKGPU
         run: PYTHONPATH="." AMD=1 MOCKGPU=1 N=256 python3 extra/gemm/amd_asm_matmul.py
       - name: Run LLVM test
diff --git a/extra/gemm/asm/cdna/gemm.py b/extra/gemm/asm/cdna/gemm.py
index 547e133e46..2b569f0d9e 100644
--- a/extra/gemm/asm/cdna/gemm.py
+++ b/extra/gemm/asm/cdna/gemm.py
@@ -42,7 +42,7 @@ def can_use_asm_gemm(a:Tensor, b:Tensor) -> bool:
   else: dname = a.device
   arch = getattr(Device[dname].renderer, "arch", "")
   if batch not in {1, 2}: return todo(f"GEMM batch size {batch}")
-  if M % TILE_M != 0 or N % TILE_N != 0 or K % TILE_K != 0:
+  if (M % TILE_M != 0 or N % TILE_N != 0 or K % TILE_K != 0) and arch == "gfx950":
     return todo(f"GEMM shape ({M},{N},{K}) not a multiple of ({TILE_M},{TILE_N},{TILE_K})")
   return True
 
diff --git a/test/testextra/test_asm_gemm.py b/test/backend/test_asm_gemm.py
similarity index 100%
rename from test/testextra/test_asm_gemm.py
rename to test/backend/test_asm_gemm.py