diff --git a/src/llama.cpp b/src/llama.cpp
index 70a7195b..9f613743 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -17823,13 +17823,16 @@ static void manage_graph_tensors(struct ggml_cgraph * cgraph, int advice, bool f
     }
 
     for (const auto & segment : merged_segments) {
+        size_t prefetch_dense = 4;
         size_t len = std::max(segment.end - segment.start, static_cast<size_t>(page_size));
         posix_madvise(reinterpret_cast<void *>(segment.start), len, advice); // hint to load into memory
         // force to prefetch data
         if (force && advice == POSIX_MADV_WILLNEED) {
             volatile char * ptr = reinterpret_cast<volatile char *>(segment.start);
-            for (size_t off = 0; off < len; off += page_size) {
-                (void)ptr[off];
+            for (size_t off = 0; off < len; off += prefetch_dense * page_size) {
+                for (size_t i = 0; i < prefetch_dense; i++) {
+                    if (off + i * page_size < len) (void)ptr[off + i * page_size];
+                }
             }
         }
     }