diff --git a/barretenberg/cpp/CMakePresets.json b/barretenberg/cpp/CMakePresets.json
index 1bab33d1186..b0bfd929265 100644
--- a/barretenberg/cpp/CMakePresets.json
+++ b/barretenberg/cpp/CMakePresets.json
@@ -184,20 +184,17 @@
       }
     },
     {
-      "name": "xray-1thread-verbose",
+      "name": "xray-verbose",
       "displayName": "Build with single-threaded XRay Profiling",
       "description": "Build with Clang and enable single-threaded LLVM XRay for profiling",
       "generator": "Unix Makefiles",
       "inherits": "clang16",
       "environment": {
-        "CFLAGS": "-fxray-instrument -fxray-instruction-threshold=100 -mllvm -finline-max-stacksize=150",
-        "CXXFLAGS": "-fxray-instrument -fxray-instruction-threshold=100 -mllvm -finline-max-stacksize=150",
-        "LDFLAGS": "-fxray-instrument -fxray-instruction-threshold=100 -mllvm -finline-max-stacksize=150"
+        "CFLAGS": "-fxray-instrument -fxray-instruction-threshold=100 -finline-max-stacksize=150",
+        "CXXFLAGS": "-fxray-instrument -fxray-instruction-threshold=100 -finline-max-stacksize=150",
+        "LDFLAGS": "-fxray-instrument -fxray-instruction-threshold=100 -finline-max-stacksize=150"
       },
-      "cacheVariables": {
-        "MULTITHREADING": "OFF"
-      },
-      "binaryDir": "build-xray-1thread-verbose"
+      "binaryDir": "build-xray-verbose"
     },
     {
       "name": "xray-1thread",
@@ -320,8 +317,8 @@
       "targets": ["barretenberg.wasm"]
     },
     {
-      "name": "xray-1thread-verbose",
-      "configurePreset": "xray-1thread-verbose",
+      "name": "xray-verbose",
+      "configurePreset": "xray-verbose",
       "inherits": "default"
     },
     {
diff --git a/barretenberg/cpp/scripts/collect_profile_information.sh b/barretenberg/cpp/scripts/collect_profile_information.sh
index 1262bb63b82..0b7d79ef8ed 100755
--- a/barretenberg/cpp/scripts/collect_profile_information.sh
+++ b/barretenberg/cpp/scripts/collect_profile_information.sh
@@ -41,6 +41,6 @@ llvm-xray-16 stack xray-log.honk_bench_main_simple.* \
   --instr_map=./bin/honk_bench_main_simple --stack-format=flame --aggregate-threads --aggregation-type=time --all-stacks \
   | node ../scripts/llvm_xray_stack_flame_corrector.js \
   | shorten_cpp_names \
-  | ../scripts/flamegraph.pl --width 2000 --fontsize 10 \
+  | ../scripts/flamegraph.pl --width 1200 --fontsize 10 \
   > xray.svg
 echo "Profiling complete, now you can do e.g. 'scp mainframe:`readlink -f xray.svg` .' on a local terminal and open the SVG in a browser."
diff --git a/barretenberg/cpp/src/barretenberg/benchmark/honk_bench/main.simple.cpp b/barretenberg/cpp/src/barretenberg/benchmark/honk_bench/main.simple.cpp
index a20b68c9e5f..b33d5a57097 100644
--- a/barretenberg/cpp/src/barretenberg/benchmark/honk_bench/main.simple.cpp
+++ b/barretenberg/cpp/src/barretenberg/benchmark/honk_bench/main.simple.cpp
@@ -46,7 +46,7 @@ template <typename Builder> void generate_sha256_test_circuit(Builder& builder,
 BBERG_INSTRUMENT BBERG_NOINLINE void sumcheck_profiling(honk::UltraProver& ext_prover)
 {
     ext_prover.construct_proof();
-    for (size_t i = 0; i < 10000; i++) {
+    for (size_t i = 0; i < 200; i++) {
         // Bench sumcheck
         ext_prover.execute_relation_check_rounds();
     }
@@ -60,7 +60,8 @@ void construct_proof_ultra() noexcept
     barretenberg::srs::init_crs_factory("../srs_db/ignition");
     // Constuct circuit and prover; don't include this part in measurement
     honk::UltraComposer::CircuitBuilder builder;
-    generate_sha256_test_circuit(builder, 100);
+    generate_sha256_test_circuit(builder, 1);
+    std::cout << "gates: " << builder.get_total_circuit_size() << std::endl;
 
     honk::UltraComposer composer;
     std::shared_ptr<honk::UltraComposer::Instance> instance = composer.create_instance(builder);
diff --git a/barretenberg/cpp/src/barretenberg/common/parallel_for_mutex_pool.cpp b/barretenberg/cpp/src/barretenberg/common/parallel_for_mutex_pool.cpp
index c8fb4ebf5ec..beb9ab35e5a 100644
--- a/barretenberg/cpp/src/barretenberg/common/parallel_for_mutex_pool.cpp
+++ b/barretenberg/cpp/src/barretenberg/common/parallel_for_mutex_pool.cpp
@@ -50,7 +50,7 @@ class ThreadPool {
     std::condition_variable complete_condition_;
     bool stop = false;
 
-    void worker_loop(size_t thread_index);
+    [[clang::xray_never_instrument]] void worker_loop(size_t thread_index);
 
     void do_iterations()
     {
diff --git a/barretenberg/cpp/src/barretenberg/honk/sumcheck/sumcheck.hpp b/barretenberg/cpp/src/barretenberg/honk/sumcheck/sumcheck.hpp
index b8348c819d2..e2f2cb335ac 100644
--- a/barretenberg/cpp/src/barretenberg/honk/sumcheck/sumcheck.hpp
+++ b/barretenberg/cpp/src/barretenberg/honk/sumcheck/sumcheck.hpp
@@ -139,12 +139,14 @@ template <typename Flavor> class SumcheckProver {
     void partially_evaluate(auto& polynomials, size_t round_size, FF round_challenge)
     {
         // after the first round, operate in place on partially_evaluated_polynomials
-        for (size_t j = 0; j < polynomials.size(); ++j) {
+        parallel_for(polynomials.size(), [&](size_t j) {
             for (size_t i = 0; i < round_size; i += 2) {
+                auto x = polynomials[j][i];
+                std::cout << polynomials[j][i] << std::endl;
                 partially_evaluated_polynomials[j][i >> 1] =
                     polynomials[j][i] + round_challenge * (polynomials[j][i + 1] - polynomials[j][i]);
             }
-        }
+        });
     };
 };