diff --git a/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
index 1352f34..5f052f7 100644
--- a/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpMatrixMultiplyKernel.cpp
@@ -52,7 +52,7 @@
 
 void NEGEMMLowpMatrixMultiplyKernel::configure(const ITensor *input0, const ITensor *input1, ITensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::S8);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
 
@@ -90,41 +90,8 @@
     INEKernel::configure(win);
 }
 
-void NEGEMMLowpMatrixMultiplyKernel::run(const Window &window, const ThreadInfo &info)
+void inline matrix_multiply_u8(Iterator &ina, Iterator &inb, Iterator &out, int width_b, size_t out_stride, const Window &window)
 {
-    ARM_COMPUTE_UNUSED(info);
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
-
-    const size_t in_b_stride = _input1->info()->strides_in_bytes()[1];
-    const size_t out_stride  = _output->info()->strides_in_bytes()[1] / _output->info()->element_size();
-
-    // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the output matrix
-    Window win_a(window);
-    win_a.set(Window::DimX, Window::Dimension(0, 0, 0));
-    win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, window.y().end() / 4, 1));
-
-    // Set step_x and step_y for matrix B. Scale by a factor of 16 the X range as the input transposed matrix A has 16 times less the columns of the output matrix
-    Window win_b;
-    // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
-    // This scenario can happen when the the matrix multiplication is used to perform a convolution operation
-    if(_slide_matrix_b)
-    {
-        win_b = window;
-    }
-    win_b.set(Window::DimX, Window::Dimension(window.x().start() / 16, window.x().end() / 16, in_b_stride));
-    win_b.set(Window::DimY, Window::Dimension(0, 0, 0));
-
-    // The step x and step y for the output matrix has been already set using in configure()
-    Iterator ina(_input0, win_a);
-    Iterator inb(_input1, win_b);
-    Iterator out(_output, window);
-
-    const int width_b = _input1->info()->dimension(0);
-
-    // The implementation assumes that the matrix A and Matrix B have been reshaped respectively with NEGEMMInterleave4x4 and NEGEMMTranspose1xW
-    // The reshaping of the matrices helps to have a cache friendly implementation and helps to avoid the data re-arrangements needed for computing 16x4 elements per iteration
-    // All the values needed for computing a single 4x4 block will be read from consecutive memory positions
     execute_window_loop(window, [&](const Coordinates & id)
     {
         const uint8_t *mtx_a0 = ina.ptr();
@@ -239,3 +206,175 @@
     },
     ina, inb, out);
 }
+
+void inline matrix_multiply_s8(Iterator &ina, Iterator &inb, Iterator &out, int width_b, size_t out_stride, const Window &window)
+{
+    // The implementation assumes that the matrix A and Matrix B have been reshaped respectively with NEGEMMInterleave4x4 and NEGEMMTranspose1xW
+    // The reshaping of the matrices helps to have a cache friendly implementation and helps to avoid the data re-arrangements needed for computing 16x4 elements per iteration
+    // All the values needed for computing a single 4x4 block will be read from consecutive memory positions
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        auto *mtx_a0 = reinterpret_cast<const int8_t *>(ina.ptr());
+        auto *mtx_b0 = reinterpret_cast<const int8_t *>(inb.ptr());
+
+        // Note: Since the input are all positives, we can use uint32_t
+        // Accumulators for the block 0
+        int32x4x4_t c0 =
+        {
+            {
+                vdupq_n_s32(0),
+                vdupq_n_s32(0),
+                vdupq_n_s32(0),
+                vdupq_n_s32(0)
+            }
+        };
+
+        // Accumulators for the block 1
+        int32x4x4_t c1 =
+        {
+            {
+                vdupq_n_s32(0),
+                vdupq_n_s32(0),
+                vdupq_n_s32(0),
+                vdupq_n_s32(0)
+            }
+        };
+
+        // Accumulators for the block 2
+        int32x4x4_t c2 =
+        {
+            {
+                vdupq_n_s32(0),
+                vdupq_n_s32(0),
+                vdupq_n_s32(0),
+                vdupq_n_s32(0)
+            }
+        };
+
+        // Accumulators for the block 3
+        int32x4x4_t c3 =
+        {
+            {
+                vdupq_n_s32(0),
+                vdupq_n_s32(0),
+                vdupq_n_s32(0),
+                vdupq_n_s32(0)
+            }
+        };
+
+        for(int k = 0; k < width_b; k += 16, mtx_a0 += 4, mtx_b0 += 16)
+        {
+            const int8x8_t  a00_s8 = vld1_s8(mtx_a0);
+            const int8x16_t b00_s8 = vld1q_s8(mtx_b0);
+
+            // Convert a00_s8 to uint16_t and get the lower part
+            const int16x4_t a00_s16 = vget_low_s16(vmovl_s8(a00_s8));
+
+            // Convert b00_s8 to int16_t
+            const int16x4x4_t b00_s16 =
+            {
+                {
+                    vget_low_s16(vmovl_s8(vget_low_s8(b00_s8))),
+                    vget_high_s16(vmovl_s8(vget_low_s8(b00_s8))),
+                    vget_low_s16(vmovl_s8(vget_high_s8(b00_s8))),
+                    vget_high_s16(vmovl_s8(vget_high_s8(b00_s8)))
+                }
+            };
+
+            // 4x4 block 0
+            c0.val[0] = vmlal_lane_s16(c0.val[0], b00_s16.val[0], a00_s16, 0);
+            c0.val[1] = vmlal_lane_s16(c0.val[1], b00_s16.val[1], a00_s16, 0);
+            c0.val[2] = vmlal_lane_s16(c0.val[2], b00_s16.val[2], a00_s16, 0);
+            c0.val[3] = vmlal_lane_s16(c0.val[3], b00_s16.val[3], a00_s16, 0);
+
+            // 4x4 block 1
+            c1.val[0] = vmlal_lane_s16(c1.val[0], b00_s16.val[0], a00_s16, 1);
+            c1.val[1] = vmlal_lane_s16(c1.val[1], b00_s16.val[1], a00_s16, 1);
+            c1.val[2] = vmlal_lane_s16(c1.val[2], b00_s16.val[2], a00_s16, 1);
+            c1.val[3] = vmlal_lane_s16(c1.val[3], b00_s16.val[3], a00_s16, 1);
+
+            // 4x4 block 2
+            c2.val[0] = vmlal_lane_s16(c2.val[0], b00_s16.val[0], a00_s16, 2);
+            c2.val[1] = vmlal_lane_s16(c2.val[1], b00_s16.val[1], a00_s16, 2);
+            c2.val[2] = vmlal_lane_s16(c2.val[2], b00_s16.val[2], a00_s16, 2);
+            c2.val[3] = vmlal_lane_s16(c2.val[3], b00_s16.val[3], a00_s16, 2);
+
+            // 4x4 block 3
+            c3.val[0] = vmlal_lane_s16(c3.val[0], b00_s16.val[0], a00_s16, 3);
+            c3.val[1] = vmlal_lane_s16(c3.val[1], b00_s16.val[1], a00_s16, 3);
+            c3.val[2] = vmlal_lane_s16(c3.val[2], b00_s16.val[2], a00_s16, 3);
+            c3.val[3] = vmlal_lane_s16(c3.val[3], b00_s16.val[3], a00_s16, 3);
+        }
+
+        auto mtx_out = reinterpret_cast<int32_t *>(out.ptr());
+        vst1q_s32(mtx_out + 0 * out_stride + 0, c0.val[0]);
+        vst1q_s32(mtx_out + 0 * out_stride + 4, c0.val[1]);
+        vst1q_s32(mtx_out + 0 * out_stride + 8, c0.val[2]);
+        vst1q_s32(mtx_out + 0 * out_stride + 12, c0.val[3]);
+        vst1q_s32(mtx_out + 1 * out_stride + 0, c1.val[0]);
+        vst1q_s32(mtx_out + 1 * out_stride + 4, c1.val[1]);
+        vst1q_s32(mtx_out + 1 * out_stride + 8, c1.val[2]);
+        vst1q_s32(mtx_out + 1 * out_stride + 12, c1.val[3]);
+        vst1q_s32(mtx_out + 2 * out_stride + 0, c2.val[0]);
+        vst1q_s32(mtx_out + 2 * out_stride + 4, c2.val[1]);
+        vst1q_s32(mtx_out + 2 * out_stride + 8, c2.val[2]);
+        vst1q_s32(mtx_out + 2 * out_stride + 12, c2.val[3]);
+        vst1q_s32(mtx_out + 3 * out_stride + 0, c3.val[0]);
+        vst1q_s32(mtx_out + 3 * out_stride + 4, c3.val[1]);
+        vst1q_s32(mtx_out + 3 * out_stride + 8, c3.val[2]);
+        vst1q_s32(mtx_out + 3 * out_stride + 12, c3.val[3]);
+    },
+    ina, inb, out);
+}
+
+void NEGEMMLowpMatrixMultiplyKernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    const size_t in_b_stride = _input1->info()->strides_in_bytes()[1];
+    const size_t out_stride  = _output->info()->strides_in_bytes()[1] / _output->info()->element_size();
+
+    // Set step_x and step_y for matrix A. Scale by a factor of 4 the Y range as the input interleaved matrix A has 4 times less the rows of the output matrix
+    Window win_a(window);
+    win_a.set(Window::DimX, Window::Dimension(0, 0, 0));
+    win_a.set(Window::DimY, Window::Dimension(window.y().start() / 4, window.y().end() / 4, 1));
+
+    // Set step_x and step_y for matrix B. Scale by a factor of 16 the X range as the input transposed matrix A has 16 times less the columns of the output matrix
+    Window win_b;
+    // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
+    // This scenario can happen when the the matrix multiplication is used to perform a convolution operation
+    if(_slide_matrix_b)
+    {
+        win_b = window;
+    }
+    win_b.set(Window::DimX, Window::Dimension(window.x().start() / 16, window.x().end() / 16, in_b_stride));
+    win_b.set(Window::DimY, Window::Dimension(0, 0, 0));
+
+    // The step x and step y for the output matrix has been already set using in configure()
+    Iterator ina(_input0, win_a);
+    Iterator inb(_input1, win_b);
+    Iterator out(_output, window);
+
+    const int width_b = _input1->info()->dimension(0);
+    switch(_input0->info()->data_type())
+    {
+        case DataType::S8:
+        {
+            matrix_multiply_s8(ina, inb, out, width_b, out_stride, window);
+            break;
+        }
+        case DataType::U8:
+        case DataType::QASYMM8:
+        {
+            matrix_multiply_u8(ina, inb, out, width_b, out_stride, window);
+            break;
+        }
+        default:
+        {
+            ARM_COMPUTE_ERROR("Not supported");
+            break;
+        }
+    }
+}
diff --git a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp
new file mode 100644
index 0000000..b75a8ab
--- /dev/null
+++ b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp
@@ -0,0 +1,129 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+#include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp"
+#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_gemm_s8_4x4.hpp"
+} // namespace arm_compute
+
+#include <arm_neon.h>
+#include <cstddef>
+#include <cstdint>
+
+// Enable only if compiled for AArch64-V8A targets
+#ifdef ARM_COMPUTE_AARCH64_V8A
+
+namespace arm_compute
+{
+void NEGEMMLowpAArch64Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::S8);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
+
+    _input0      = input0;
+    _input1      = input1;
+    _output      = output;
+    _workspace   = workspace;
+    _alpha       = alpha;
+    _beta        = beta;
+    _transform_0 = transform_0;
+    _transform_1 = transform_1;
+
+    // Configure kernel window
+    Window win = calculate_max_window(*output->info());
+
+    AccessWindowRectangle output_access(output->info(), 0, 0, 4, 4);
+
+    const int input0_access_end = ceil_to_multiple(input0->info()->tensor_shape().x(), 4);
+    const int input1_access_end = ceil_to_multiple(input1->info()->tensor_shape().x(), 4);
+
+    update_window_and_padding(win,
+                              AccessWindowStatic(input0->info(), 0, 0, input0_access_end, input0->info()->tensor_shape().y()),
+                              AccessWindowStatic(input1->info(), 0, 0, input1_access_end, input1->info()->tensor_shape().y()),
+                              output_access);
+
+    INEKernel::configure(win);
+}
+
+void NEGEMMLowpAArch64Kernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
+
+    const int lda = _input0->info()->strides_in_bytes().y();
+    const int ldb = _input1->info()->strides_in_bytes().y();
+    const int ldc = _output->info()->strides_in_bytes().y() / sizeof(int32_t);
+
+    const auto in1_ptr = reinterpret_cast<const int8_t *>(_input1->buffer());
+
+    const int M = std::min(_output->info()->tensor_shape().y(), static_cast<size_t>(window.y().end())) - window.y().start();
+    const int N = _output->info()->tensor_shape().x();
+    const int K = _input0->info()->tensor_shape().x();
+
+    // Only iterate over batches
+    Window win(window);
+    win.set(0, Window::Dimension(0, 1, 1));
+    win.set(1, Window::Dimension(0, 1, 1));
+
+    Iterator in0(_input0, window);
+    Iterator out(_output, window);
+
+    GemmInterleaved<gemm_s8_4x4, int8_t, int32_t> gemm(&info.cpu_info, M, N, K, !_transform_1, !_transform_1);
+
+    constexpr size_t alignment      = 4096;
+    const size_t     offset         = (gemm.get_working_size() + alignment - 1) * info.thread_id;
+    void            *workspace      = _workspace->buffer() + offset;
+    size_t           workspace_size = _workspace->info()->total_size();
+
+    if(support::cpp11::align(alignment, gemm.get_working_size(), workspace, workspace_size) == nullptr)
+    {
+        ARM_COMPUTE_ERROR("Not enough space to align buffer!");
+    }
+
+    execute_window_loop(win, [&](const Coordinates & id)
+    {
+        gemm.execute(reinterpret_cast<const int8_t *>(in0.ptr()), lda,
+                     reinterpret_cast<const int8_t *>(in1_ptr), ldb,
+                     reinterpret_cast<int32_t *>(out.ptr()), ldc,
+                     _alpha, _beta, workspace);
+    },
+    in0, out);
+}
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_AARCH64_V8A */