diff --git a/qdp_project/src/Benchmark.cpp b/qdp_project/src/Benchmark.cpp
index 7c2f76b..a7bc547 100644
--- a/qdp_project/src/Benchmark.cpp
+++ b/qdp_project/src/Benchmark.cpp
@@ -19,7 +19,7 @@
 
 #include "BenchmarkHelpers.cpp"
 
-#define MODE_PREFETCH
+#define MODE_HBM
 
 ////////////////////////////////
 /// BENCHMARK SETUP
@@ -29,7 +29,7 @@ constexpr uint32_t WARMUP_ITERATION_COUNT = 5;
 constexpr uint32_t ITERATION_COUNT = 5;
 
 #ifdef MODE_PREFETCH
-constexpr size_t CHUNK_SIZE_B = 64_MiB;
+constexpr size_t CHUNK_SIZE_B = 128_MiB;
 constexpr uint32_t GROUP_COUNT = 32;
 constexpr uint32_t TC_SCANA = 1;
 constexpr uint32_t TC_SCANB = 1;
@@ -73,7 +73,6 @@ constexpr size_t MASK_STEP_SIZE = CHUNK_SIZE_ELEMENTS / MASK_ELEMENT_SIZE;
 
 static_assert(RUN_COUNT > 0);
 static_assert(TC_SCANB <= TC_AGGRJ);
-static_assert(TC_AGGRJ % TC_SCANB == 0);
 static_assert(WL_SIZE_B % 16 == 0);
 static_assert(CHUNK_SIZE_B % 16 == 0);
 
@@ -108,7 +107,7 @@ inline uint16_t* get_mask(uint16_t* base, const size_t chunk_index, const size_t
 }
 
 void scan_b(size_t gid, size_t tid) {
-    constexpr size_t split =  TC_AGGRJ / TC_SCANB;
+    constexpr size_t split =  TC_AGGRJ / (TC_SCANB == 0 ? 1 : TC_SCANB);
     const size_t start = tid * split;
     const size_t end = start + split;