Merge pull request #5344 from Akshay-Venkatesh/topic/update-cuda-ipc-iface

UCT/CUDA: Update cuda-ipc bandwidth for Ampere

Author: yosefe

src/uct/cuda/base/cuda_iface.h

@@ -76,8 +76,9 @@
 
 
 typedef enum uct_cuda_base_gen {
-    UCT_CUDA_BASE_GEN_PASCAL = 6,
-    UCT_CUDA_BASE_GEN_VOLTA  = 7
+    UCT_CUDA_BASE_GEN_P100 = 6,
+    UCT_CUDA_BASE_GEN_V100 = 7,
+    UCT_CUDA_BASE_GEN_A100 = 8
 } uct_cuda_base_gen_t;
 
 

src/uct/cuda/cuda_ipc/cuda_ipc_iface.c

@@ -100,12 +100,18 @@ static double uct_cuda_ipc_iface_get_bw()
         return 0;
     }
 
-    /* TODO: Detect nvswitch */
+    /*
+     * TODO: Detect nvswitch
+     * TODO: Not reporting peak unidirectional bandwidth to avoid dropping other
+     *       transports like cma/knem/ib in rma_bw_lanes
+     */
     switch (major_version) {
-    case UCT_CUDA_BASE_GEN_PASCAL:
+    case UCT_CUDA_BASE_GEN_P100:
         return 20000.0 * UCS_MBYTE;
-    case UCT_CUDA_BASE_GEN_VOLTA:
+    case UCT_CUDA_BASE_GEN_V100:
         return 25000.0 * UCS_MBYTE;
+    case UCT_CUDA_BASE_GEN_A100:
+        return 30000.0 * UCS_MBYTE;
     default:
         return 6911.0  * UCS_MBYTE;
     }