diff --git a/scc b/scc
index 3d5f733..a21316a 160000
--- a/scc
+++ b/scc
@@ -1 +1 @@
-Subproject commit 3d5f733ed40ffdbd8f9f788a5a5075fe0797498c
+Subproject commit a21316a97fff4953a3c08d32f6b97cf6d11d441e
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 173dc7f..f50c669 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -1,6 +1,7 @@
 add_subdirectory(io-redirector)
 add_subdirectory(ordered_semaphore)
 add_subdirectory(axi4_pin_level)
+add_subdirectory(ace_pin_level)
 add_subdirectory(configuration)
 if(FULL_TEST_SUITE)
 	add_subdirectory(sim_performance)
diff --git a/tests/ace_pin_level/CMakeLists.txt b/tests/ace_pin_level/CMakeLists.txt
new file mode 100644
index 0000000..78d5cf1
--- /dev/null
+++ b/tests/ace_pin_level/CMakeLists.txt
@@ -0,0 +1,9 @@
+project (ace_pin_level)
+
+add_executable(${PROJECT_NAME} 
+	ace_narrow_burst_test.cpp
+	${test_util_SOURCE_DIR}/sc_main.cpp
+)
+target_link_libraries (${PROJECT_NAME} PUBLIC test_util)
+
+catch_discover_tests(${PROJECT_NAME})
diff --git a/tests/ace_pin_level/ace_narrow_burst_test.cpp b/tests/ace_pin_level/ace_narrow_burst_test.cpp
new file mode 100644
index 0000000..5800a76
--- /dev/null
+++ b/tests/ace_pin_level/ace_narrow_burst_test.cpp
@@ -0,0 +1,242 @@
+
+#include "testbench.h"
+#include <factory.h>
+#include <tlm/scc/tlm_gp_shared.h>
+#undef CHECK
+#include <catch2/catch_all.hpp>
+#include <unordered_map>
+
+using namespace sc_core;
+
+factory::add<testbench> tb;
+
+bool operator==(tlm::tlm_generic_payload const& a, tlm::tlm_generic_payload const& b){
+    auto ret = true;
+    ret &= a.get_command() == b.get_command();
+    ret &= a.get_address() == b.get_address();
+    ret &= a.get_data_length() == b.get_data_length();
+    for(auto i=0u; i<a.get_data_length(); ++i)
+        ret &= a.get_data_ptr()[i] == b.get_data_ptr()[i];
+    //    if(a.get_byte_enable_ptr() && b.get_byte_enable_ptr()) {
+    //        ret &= a.get_byte_enable_length() == b.get_byte_enable_length();
+    //        for(auto i=0u; i<a.get_byte_enable_length(); ++i)
+    //            ret &= a.get_byte_enable_ptr()[i] == b.get_byte_enable_ptr()[i];
+    //    }
+    ret &= a.get_command() == b.get_command();
+    //if(!ret) SCCWARN()<<"Comparison failed: "<<a<<" and "<<b;
+    return ret;
+}
+
+template<typename bus_cfg>
+tlm::tlm_generic_payload* prepare_trans(uint64_t start_address, unsigned addr_incr, unsigned len, unsigned width, unsigned id) {
+    auto trans = tlm::scc::tlm_mm<>::get().allocate<axi::axi4_extension>(len);
+    trans->set_address(start_address);
+    tlm::scc::setId(*trans, id);
+    auto ext = trans->get_extension<axi::axi4_extension>();
+    trans->set_data_length(len);
+    trans->set_streaming_width(len);
+    ext->set_size(scc::ilog2(width));
+    sc_assert(len < (bus_cfg::BUSWIDTH / 8) || len % (bus_cfg::BUSWIDTH / 8) == 0);
+    auto length = (len * 8 - 1) / (8*width);
+    if(width==(bus_cfg::BUSWIDTH / 8) && start_address % (bus_cfg::BUSWIDTH / 8))
+        length++;
+    ext->set_length(length);
+    // ext->set_burst(len * 8 > bus_cfg::buswidth ? axi::burst_e::INCR : axi::burst_e::FIXED);
+    ext->set_burst(axi::burst_e::INCR);
+    ext->set_id(id);
+    return trans;
+}
+
+template<typename bus_cfg>
+tlm::tlm_generic_payload* prepare_trans_ace(uint64_t start_address, unsigned addr_incr, unsigned len, unsigned width, unsigned id) {
+    auto trans = tlm::scc::tlm_mm<>::get().allocate<axi::ace_extension>(len);
+    trans->set_address(start_address);
+    tlm::scc::setId(*trans, id);
+    auto ext = trans->get_extension<axi::ace_extension>();
+    trans->set_data_length(len);
+    trans->set_streaming_width(len);
+    ext->set_size(scc::ilog2(width));
+    sc_assert(len < (bus_cfg::BUSWIDTH / 8) || len % (bus_cfg::BUSWIDTH / 8) == 0);
+    auto length = (len * 8 - 1) / (8*width);
+    if(width==(bus_cfg::BUSWIDTH / 8) && start_address % (bus_cfg::BUSWIDTH / 8))
+        length++;
+    ext->set_length(length);
+    // ext->set_burst(len * 8 > bus_cfg::buswidth ? axi::burst_e::INCR : axi::burst_e::FIXED);
+    ext->set_burst(axi::burst_e::INCR);
+    ext->set_id(id);
+    ext->set_snoop(axi::snoop_e::READ_SHARED); // set it so that is_data_less return true???
+    return trans;
+}
+
+inline void randomize(tlm::tlm_generic_payload& gp) {
+    static uint8_t  req_cnt{0};
+    auto addr = gp.get_address();
+    uint8_t const* src = reinterpret_cast<uint8_t const*>(&addr);
+    for(size_t i = 0; i < gp.get_data_length(); ++i) {
+        *(gp.get_data_ptr() + i) = i % 2 ? i : req_cnt;
+    }
+    req_cnt++;
+}
+
+template<typename STATE>
+unsigned run_scenario(STATE& state){
+    auto& dut = factory::get<testbench>();
+
+    dut.axi_tgt_pe.set_operation_cb([&state](axi::axi_protocol_types::tlm_payload_type& trans) -> unsigned {
+        auto id = axi::get_axi_id(trans);
+        if(trans.is_read()) {
+            for(size_t i = 0; i < trans.get_data_length(); ++i) {
+                *(trans.get_data_ptr() + i) = i % 2 ? 123 : (state.resp_cnt+128);
+            }
+            state.read_tx[id].second.emplace_back(&trans);
+        }
+        if(trans.is_write())
+            state.write_tx[id].second.emplace_back(&trans);
+        SCCDEBUG(__FUNCTION__)<<"RX: "<<trans;
+        state.resp_cnt++;
+        return 0;
+    });
+
+
+
+    dut.rst.write(false);
+    sc_start(state.ResetCycles*dut.clk.period());
+    dut.rst.write(true);
+    sc_start(dut.clk.period());
+
+    auto run1 = sc_spawn([&dut, &state](){
+        unsigned int StartAddr{0x0};
+        for(int i = 0; i < state.NumberOfIterations; ++i) {
+            // tlm::scc::tlm_gp_shared_ptr trans = prepare_trans<testbench::bus_cfg>(StartAddr, 4, state.BurstLengthByte, state.BurstSizeBytes, 1);
+            tlm::scc::tlm_gp_shared_ptr trans = prepare_trans_ace<testbench::bus_cfg>(StartAddr, 4, state.BurstLengthByte, state.BurstSizeBytes, 1);
+            trans->set_command(tlm::TLM_READ_COMMAND);
+            SCCDEBUG(__FUNCTION__) << "run1, iteration " << i <<" TX: "<<*trans;
+            dut.intor_pe.transport(*trans, false);
+            state.read_tx[axi::get_axi_id(*trans)].first.emplace_back(trans);
+            StartAddr += state.BurstSizeBytes;
+        }
+    });
+/*
+    auto run2 = sc_spawn([&dut, &state](){
+        unsigned int StartAddr{0x2000};
+        for(int i = 0; i < state.NumberOfIterations; ++i) {
+            tlm::scc::tlm_gp_shared_ptr trans = prepare_trans_ace<testbench::bus_cfg>(StartAddr, 4, state.BurstLengthByte, state.BurstSizeBytes, 2);
+            trans->set_command(tlm::TLM_WRITE_COMMAND);
+            randomize(*trans);
+            SCCDEBUG(__FUNCTION__) << "run2, iteration " << i <<" TX: "<<*trans;
+            dut.intor_pe.transport(*trans, false);
+            state.write_tx[axi::get_axi_id(*trans)].first.emplace_back(trans);
+            StartAddr += state.BurstSizeBytes;
+        }
+    });
+    auto run3 = sc_spawn([&dut, &state](){
+        unsigned int StartAddr{0x1000};
+        for(int i = 0; i < state.NumberOfIterations; ++i) {
+            tlm::scc::tlm_gp_shared_ptr trans = prepare_trans_ace<testbench::bus_cfg>(StartAddr, 4, state.BurstLengthByte, state.BurstSizeBytes, 3);
+            trans->set_command(tlm::TLM_READ_COMMAND);
+            SCCDEBUG(__FUNCTION__) << "run3, iteration " << i <<" TX: "<<*trans;
+            dut.intor_pe.transport(*trans, false);
+            state.read_tx[axi::get_axi_id(*trans)].first.emplace_back(trans);
+            StartAddr += state.BurstSizeBytes;
+        }
+    });
+    auto run4 = sc_spawn([&dut, &state](){
+        unsigned int StartAddr{0x3000};
+        for(int i = 0; i < state.NumberOfIterations; ++i) {
+            tlm::scc::tlm_gp_shared_ptr trans = prepare_trans_ace<testbench::bus_cfg>(StartAddr, 4, state.BurstLengthByte, state.BurstSizeBytes, 4);
+            trans->set_command(tlm::TLM_WRITE_COMMAND);
+            randomize(*trans);
+            SCCDEBUG(__FUNCTION__) << "run4, iteration " << i <<" TX: "<<*trans;
+            dut.intor_pe.transport(*trans, false);
+            state.write_tx[axi::get_axi_id(*trans)].first.emplace_back(trans);
+            StartAddr += state.BurstSizeBytes;
+        }
+    });
+*/
+    unsigned cycles{0};
+ //   while(cycles<1000 && !(run1.terminated() && run2.terminated() && run3.terminated() && run4.terminated())){
+   while(cycles<1000 && !(run1.terminated())){
+        sc_start(10 * dut.clk.period());
+        cycles+=10;
+    }
+    return cycles;
+}
+
+TEST_CASE("ace_burst_alignment", "[AXI][pin-level]") {
+    struct {
+        unsigned int ResetCycles{4};
+ //       unsigned int BurstLengthByte{16};
+        unsigned int BurstLengthByte{32};
+        unsigned int BurstSizeBytes{8};
+//        unsigned int NumberOfIterations{8};
+        unsigned int NumberOfIterations{1};
+        std::unordered_map<unsigned, std::pair<std::vector<tlm::scc::tlm_gp_shared_ptr>, std::vector<tlm::scc::tlm_gp_shared_ptr>>> read_tx;
+        std::unordered_map<unsigned, std::pair<std::vector<tlm::scc::tlm_gp_shared_ptr>, std::vector<tlm::scc::tlm_gp_shared_ptr>>> write_tx;
+        unsigned resp_cnt{0};
+    } state;
+
+    auto cycles = run_scenario(state);
+
+    REQUIRE(cycles<1000);
+    REQUIRE(sc_report_handler::get_count(SC_ERROR) == 0);
+    REQUIRE(sc_report_handler::get_count(SC_WARNING) == 0);
+
+    REQUIRE(state.resp_cnt==4*state.NumberOfIterations);
+    for(auto& e: state.write_tx) {
+        auto const& send_tx = e.second.first;
+        auto const& recv_tx = e.second.second;
+        REQUIRE(send_tx.size() == recv_tx.size());
+        for(auto i = 0; i<send_tx.size(); ++i) {
+            REQUIRE(send_tx[i]->get_response_status() == tlm::TLM_OK_RESPONSE);
+            CHECK(*send_tx[i] == *recv_tx[i]);
+        }
+
+    }
+    for(auto& e: state.read_tx) {
+        auto const& send_tx = e.second.first;
+        auto const& recv_tx = e.second.second;
+        REQUIRE(send_tx.size() == recv_tx.size());
+        for(auto i = 0; i<send_tx.size(); ++i){
+            REQUIRE(send_tx[i]->get_response_status() == tlm::TLM_OK_RESPONSE);
+            CHECK(*send_tx[i] == *recv_tx[i]);
+        }
+    }
+}
+
+TEST_CASE("axi4_narrow_burst", "[AXI][pin-level]") {
+    struct {
+        unsigned int ResetCycles{4};
+        unsigned int BurstLengthByte{32};
+  //      unsigned int BurstLengthByte{16};
+        unsigned int BurstSizeBytes{4};
+ //       unsigned int NumberOfIterations{8};
+        unsigned int NumberOfIterations{1};
+        std::unordered_map<unsigned, std::pair<std::vector<tlm::scc::tlm_gp_shared_ptr>, std::vector<tlm::scc::tlm_gp_shared_ptr>>> read_tx;
+        std::unordered_map<unsigned, std::pair<std::vector<tlm::scc::tlm_gp_shared_ptr>, std::vector<tlm::scc::tlm_gp_shared_ptr>>> write_tx;
+        unsigned resp_cnt{0};
+    } state;
+
+    auto cycles = run_scenario(state);
+
+    REQUIRE(cycles<1000);
+    REQUIRE(sc_report_handler::get_count(SC_ERROR) == 0);
+    REQUIRE(sc_report_handler::get_count(SC_WARNING) == 0);
+
+    REQUIRE(state.resp_cnt==4*state.NumberOfIterations);
+    for(auto& e: state.write_tx) {
+        auto const& send_tx = e.second.first;
+        auto const& recv_tx = e.second.second;
+        REQUIRE(send_tx.size() == recv_tx.size());
+        for(auto i = 0; i<send_tx.size(); ++i)
+            CHECK(*send_tx[i] == *recv_tx[i]);
+
+    }
+    for(auto& e: state.read_tx) {
+        auto const& send_tx = e.second.first;
+        auto const& recv_tx = e.second.second;
+        REQUIRE(send_tx.size() == recv_tx.size());
+        for(auto i = 0; i<send_tx.size(); ++i)
+            CHECK(*send_tx[i] == *recv_tx[i]);
+
+    }
+}
diff --git a/tests/ace_pin_level/testbench.h b/tests/ace_pin_level/testbench.h
new file mode 100644
index 0000000..faf5fa0
--- /dev/null
+++ b/tests/ace_pin_level/testbench.h
@@ -0,0 +1,83 @@
+#ifndef _TESTBENCH_H_
+#define _TESTBENCH_H_
+
+#include <axi/pe/axi_initiator.h>
+#include <axi/pe/simple_target.h>
+#include <axi/pe/simple_ace_target.h>
+#include <axi/pin/ace_initiator.h>
+#include <axi/pin/ace_target.h>
+#include <axi/scv/recorder_modules.h>
+#include <scc.h>
+
+using namespace sc_core;
+using namespace axi;
+using namespace axi::pe;
+
+class testbench : public sc_core::sc_module {
+public:
+    using bus_cfg = axi::ace_cfg</*BUSWIDTH=*/64, /*ADDRWIDTH=*/32, /*IDWIDTH=*/4, /*USERWIDTH=*/1>;
+
+    sc_core::sc_time clk_period{10, sc_core::SC_NS};
+    sc_core::sc_clock clk{"clk", clk_period, 0.5, sc_core::SC_ZERO_TIME, true};
+    sc_core::sc_signal<bool> rst{"rst"};
+    // initiator side
+    axi::ace_initiator_socket<bus_cfg::BUSWIDTH> intor{"ace_intor"};
+
+    axi::pin::ace_initiator<bus_cfg> intor_bfm{"ace_intor_bfm"};
+    // signal accurate bus
+    axi::aw_ch_ace<bus_cfg, axi::signal_types> aw;
+    axi::wdata_ch_ace<bus_cfg, axi::signal_types> wdata;
+    axi::b_ch_ace<bus_cfg, axi::signal_types> b;
+    axi::ar_ch_ace<bus_cfg, axi::signal_types> ar;
+    axi::rresp_ch_ace<bus_cfg, axi::signal_types> rresp;
+
+    axi::pin::ace_target<bus_cfg> tgt_bfm{"ace_tgt_bfm"};
+    // target side
+    axi::ace_target_socket<bus_cfg::BUSWIDTH> tgt_ace{"tgt_ace"};
+    axi::axi_target_socket<bus_cfg::BUSWIDTH> tgt_axi{"tgt_axi"};
+    // engines
+    axi::pe::ace_initiator<bus_cfg::BUSWIDTH> intor_pe;
+    axi::pe::simple_target<bus_cfg::BUSWIDTH> axi_tgt_pe;
+    axi::pe::simple_ace_target<bus_cfg::BUSWIDTH> ace_tgt_pe;
+
+public:
+    SC_HAS_PROCESS(testbench);
+    testbench(): testbench("testbench") {}
+    testbench(sc_core::sc_module_name nm)
+    : sc_core::sc_module(nm)
+    , intor_pe("ace_intor_pe", intor)
+    , ace_tgt_pe("ace_tgt_pe", tgt_ace)
+    , axi_tgt_pe("axi_tgt_pe", tgt_axi) {
+        intor_pe.clk_i(clk);
+        intor_bfm.clk_i(clk);
+        tgt_bfm.clk_i(clk);
+        axi_tgt_pe.clk_i(clk);
+        ace_tgt_pe.clk_i(clk);
+        // pe socket to recorder
+        intor(intor_bfm.tsckt);
+
+        // bfm to signals
+        intor_bfm.bind_aw(aw);
+        intor_bfm.bind_w(wdata);
+        intor_bfm.bind_b(b);
+        intor_bfm.bind_ar(ar);
+        intor_bfm.bind_r(rresp);
+        // signals to bfm
+        tgt_bfm.bind_aw(aw);
+        tgt_bfm.bind_w(wdata);
+        tgt_bfm.bind_b(b);
+        tgt_bfm.bind_ar(ar);
+        tgt_bfm.bind_r(rresp);
+        // bfm to recorder
+        tgt_bfm.isckt(tgt_ace);
+
+        ace_tgt_pe.isckt_axi(tgt_axi);
+
+    }
+
+
+    void run1() {
+    }
+};
+
+#endif // _TESTBENCH_H_