CUDA: correct the lowest Maxwell supported by CUDA 12 (#11984)

* CUDA: correct the lowest Maxwell supported by CUDA 12 --------- Co-authored-by: Johannes Gäßler <[email protected]>
ggml-org · Feb 21, 2025 · ecc8e3a · ecc8e3a
1 parent 0b3863f
commit ecc8e3a
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -7,7 +7,7 @@ if (CUDAToolkit_FOUND)
 
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
         # native == GPUs available at build time
-        # 52     == Maxwell, lowest CUDA 12 standard
+        # 50     == Maxwell, lowest CUDA 12 standard
         # 60     == P100, FP16 CUDA intrinsics
         # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
         # 70     == V100, FP16 tensor cores
@@ -17,7 +17,7 @@ if (CUDAToolkit_FOUND)
         elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75;80")
         else()
-            set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75;80")
+            set(CMAKE_CUDA_ARCHITECTURES "50;61;70;75;80")
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")