diff --git a/include/my_cpu.h b/include/my_cpu.h
index 028dc05bd084a..66205de99a408 100644
--- a/include/my_cpu.h
+++ b/include/my_cpu.h
@@ -110,6 +110,25 @@ static inline void MY_RELAX_CPU(void)
 }
 
 
+/*
+  Hint the CPU to start loading a cache line for reading.
+
+  Like YieldProcessor() above, PreFetchCacheLine() is architecture-independent:
+  winnt.h picks the instruction per target, unlike _mm_prefetch().
+  clang-cl defines __clang__ but not __GNUC__, so it must take the first branch.
+*/
+static inline void my_prefetch(const void *addr)
+{
+#if defined(__GNUC__) || defined(__clang__)
+  __builtin_prefetch(addr, 0, 3);
+#elif defined(_WIN32)
+  PreFetchCacheLine(PF_TEMPORAL_LEVEL_1, addr);
+#else
+  (void) addr;
+#endif
+}
+
+
 #ifdef HAVE_PAUSE_INSTRUCTION
 # ifdef __cplusplus
 extern "C" {
diff --git a/sql/vector_mhnsw.cc b/sql/vector_mhnsw.cc
index c480c36c7e7ad..d42461cae6620 100644
--- a/sql/vector_mhnsw.cc
+++ b/sql/vector_mhnsw.cc
@@ -22,6 +22,7 @@
 #include "vector_mhnsw.h"
 #include <scope.h>
 #include <my_atomic_wrapper.h>
+#include <my_cpu.h>                            // my_prefetch()
 #include "bloom_filters.h"
 
 // distance can be a little bit < 0 because of fast math
@@ -1347,6 +1348,19 @@ static int search_layer(MHNSW_param *p, const FVector *target, float threshold,
       if (res == 0xff)
         continue;
 
+      // A node and its vector share one allocation. Prefetch unseen nodes before
+      // computing distances so later loads can overlap with work on earlier ones.
+      for (size_t i= 0; i < 8; i++)
+      {
+        FVectorNode *link= links[i];
+        if (!(res & (1 << i)) && link)
+        {
+          my_prefetch(link);
+          my_prefetch(reinterpret_cast<const char*>(link)
+                      + CPU_LEVEL1_DCACHE_LINESIZE);
+        }
+      }
+
       for (size_t i= 0; i < 8; i++)
       {
         if (res & (1 << i))