diff --git a/sw/dnn/gemm/data/datagen.py b/sw/dnn/gemm/data/datagen.py
deleted file mode 100755
index 17f4df0496..0000000000
--- a/sw/dnn/gemm/data/datagen.py
+++ /dev/null
@@ -1,157 +0,0 @@
-#!/usr/bin/env python3
-# Copyright 2023 ETH Zurich and University of Bologna.
-# Licensed under the Apache License, Version 2.0, see LICENSE for details.
-# SPDX-License-Identifier: Apache-2.0
-#
-# Tim Fischer <fischeti@iis.ee.ethz.ch>
-# Viviane Potocnik <vivianep@iis.ee.ethz.ch>
-# Luca Colagrande <colluca@iis.ee.ethz.ch>
-
-import argparse
-import pathlib
-import json5
-import sys
-import os
-import torch
-
-sys.path.append(os.path.join(os.path.dirname(__file__), "../../../../util/sim/"))
-import data_utils  # noqa: E402
-from data_utils import emit_license, \
-                       format_struct_definition, format_array_definition, \
-                       format_array_declaration, format_ifdef_wrapper  # noqa: E402
-
-torch.manual_seed(42)
-
-# AXI splits bursts crossing 4KB address boundaries. To minimize
-# the occurrence of these splits the data should be aligned to 4KB
-BURST_ALIGNMENT = 4096
-
-
-def rand_data_generator(shape, prec, alt=False):
-    if prec == 'FP64':
-        return torch.randn(shape, requires_grad=False, dtype=torch.float64), {}
-    elif prec == 'FP32':
-        return torch.randn(shape, requires_grad=False, dtype=torch.float32), {}
-    elif prec == 'FP16':
-        if alt:
-            return torch.randn(shape, requires_grad=False, dtype=torch.bfloat16), {}
-        else:
-            return torch.randn(shape, requires_grad=False, dtype=torch.float16), {}
-    elif prec == 'FP8':
-        sign = torch.randint(0, 2, shape,
-                             requires_grad=False, dtype=torch.uint8)  # -1 or 1
-        exponent = torch.randint(0, 16, shape,
-                                 requires_grad=False, dtype=torch.uint8)  # < 0b01111
-        mantissa = torch.randint(0, 4, shape,
-                                 requires_grad=False, dtype=torch.uint8)  # can be arbitrary
-        bits = {'sign': sign, 'exponent': exponent, 'mantissa': mantissa}
-        # TODO: not actually correct
-        sign_val = (-1.0)**sign.double()
-        exp_val = (2.0**(exponent.double()-15.0))
-        man_val = (1.0 + mantissa.double() / (2**2))
-        val = sign_val*exp_val*man_val
-        return val, bits
-
-
-def golden_model(alpha, A, B, C):
-    return alpha * C + torch.matmul(A, B)
-
-
-def emit_header(**kwargs):
-
-    M = kwargs['M']
-    N = kwargs['N']
-    K = kwargs['K']
-    alpha = kwargs['alpha']
-    expand = kwargs['expand']
-    transpose_A = kwargs['transpose_A']
-    transpose_B = kwargs['transpose_B']
-    prec = kwargs['prec']
-
-    print(prec)
-    mat_A, bits_A = rand_data_generator((M, K), prec)
-    mat_B, bits_B = rand_data_generator((K, N), prec)
-    mat_C, bits_C = rand_data_generator((M, N), prec)
-
-    result = golden_model(alpha, mat_A, mat_B, mat_C)
-
-    if transpose_A:
-        mat_A = mat_A.T
-    if transpose_B:
-        mat_B = mat_B.T
-
-    ctype = data_utils.ctype_from_precision_t(prec)
-
-    A_uid = 'A'
-    B_uid = 'B'
-    C_uid = 'C'
-
-    layer_cfg = {
-        'M': M,
-        'N': N,
-        'K': K,
-        'TA': int(transpose_A),
-        'TB': int(transpose_B),
-        'ALPHA': alpha,
-        'expand': expand,
-        'dtype': prec,
-        'A': A_uid,
-        'B': B_uid,
-        'C': C_uid
-    }
-
-    data_str = [emit_license()]
-    # Array forward declarations
-    data_str += [format_array_declaration(ctype, A_uid, mat_A.shape)]
-    data_str += [format_array_declaration(ctype, B_uid, mat_B.shape)]
-    data_str += [format_array_declaration(ctype, C_uid, mat_C.shape)]
-    # Layer struct
-    data_str += [format_struct_definition('gemm_layer_t', 'layer', layer_cfg)]
-    # Array definitions
-    if prec == 'FP8':
-        data_str += [format_array_definition(ctype, A_uid, bits_A)]
-        data_str += [format_array_definition(ctype, B_uid, bits_B)]
-        data_str += [format_array_definition(ctype, C_uid, bits_C)]
-    else:
-        data_str += [format_array_definition(ctype, A_uid, mat_A)]
-        data_str += [format_array_definition(ctype, B_uid, mat_B)]
-        data_str += [format_array_definition(ctype, C_uid, mat_C)]
-    # Golden results for BIST
-    result_def = format_array_definition(ctype, 'checksum', torch.sum(result, dim=-1))
-    data_str += [format_ifdef_wrapper('BIST', result_def)]
-    data_str = '\n\n'.join(data_str)
-
-    return data_str
-
-
-def main():
-
-    parser = argparse.ArgumentParser(description='Generate data for layernorm kernel')
-    parser.add_argument(
-        "-c", "--cfg",
-        type=pathlib.Path,
-        required=True,
-        help='Select param config file kernel'
-    )
-    parser.add_argument(
-        '--section',
-        type=str,
-        help='Section to store matrices in')
-    parser.add_argument(
-        'output',
-        type=pathlib.Path,
-        help='Path of the output header file')
-    args = parser.parse_args()
-
-    # Load param config file
-    with args.cfg.open() as f:
-        param = json5.loads(f.read())
-    param['section'] = args.section
-
-    # Emit header file
-    with open(args.output, 'w') as f:
-        f.write(emit_header(**param))
-
-
-if __name__ == '__main__':
-    main()
diff --git a/sw/dnn/gemm/data/params.json b/sw/dnn/gemm/data/params.json
deleted file mode 100644
index 11badbf202..0000000000
--- a/sw/dnn/gemm/data/params.json
+++ /dev/null
@@ -1,14 +0,0 @@
-// Copyright 2020 ETH Zurich and University of Bologna.
-// Solderpad Hardware License, Version 0.51, see LICENSE for details.
-// SPDX-License-Identifier: SHL-0.51
-
-{
-    M: 16,
-    N: 16,
-    K: 16,
-    alpha: 0,
-    transpose_A: false,
-    transpose_B: true,
-    prec: "FP32",
-    expand: 0
-}
diff --git a/sw/dnn/gemm/src/gemm.h b/sw/dnn/gemm/src/gemm.h
deleted file mode 100644
index cf2b2949e0..0000000000
--- a/sw/dnn/gemm/src/gemm.h
+++ /dev/null
@@ -1,65 +0,0 @@
-// Copyright 2020 ETH Zurich and University of Bologna.
-// Licensed under the Apache License, Version 2.0, see LICENSE for details.
-// SPDX-License-Identifier: Apache-2.0
-
-#pragma once
-
-#include <stdint.h>
-#include "blas.h"
-
-/**
- * @struct gemm_layer_struct
- * @brief This structure contains all parameters necessary for GEMM.
- * @var gemm_layer_struct::M
- * Dimension of matrix product MxK * KxN
- * @var gemm_layer_struct::M_p
- * M divided by number of compute cores
- * @var gemm_layer_struct::N
- * Dimension of matrix product MxK * KxN
- * @var gemm_layer_struct::K
- * Dimension of matrix product MxK * KxN
- * @var gemm_layer_struct::TA
- * Transpose matrix A
- * @var gemm_layer_struct::TB
- * Transpose matrix B
- * @var gemm_layer_struct::TILE_M
- * Tile factor across M dimension
- * @var gemm_layer_struct::TILE_N
- * Tile factor across N dimension
- * @var gemm_layer_struct::TILE_K
- * Tile factor across K dimension
- * @var gemm_layer_struct::A
- * Pointer to matrix A
- * @var gemm_layer_struct::B
- * Pointer to matrix B
- * @var gemm_layer_struct::C
- * Pointer to matrix C
- * @var gemm_layer_struct::ALPHA
- * constant factor: A * B + ALPHA * C
- * @var gemm_layer_struct::dtype
- * Precision of GEMM
- * @var gemm_layer_struct::expand
- * Use expanding DOTP instructions
- */
-typedef struct gemm_layer_struct {
-    uint32_t M;
-    uint32_t M_p;
-    uint32_t N;
-    uint32_t K;
-
-    uint32_t TA;
-    uint32_t TB;
-
-    uint32_t TILE_M;
-    uint32_t TILE_N;
-    uint32_t TILE_K;
-
-    void *A;
-    void *B;
-    void *C;
-
-    uint32_t ALPHA;
-
-    precision_t dtype;
-    uint32_t expand;
-} gemm_layer_t;
diff --git a/sw/dnn/gemm/src/main.c b/sw/dnn/gemm/src/main.c
deleted file mode 100644
index 62ca12cf29..0000000000
--- a/sw/dnn/gemm/src/main.c
+++ /dev/null
@@ -1,207 +0,0 @@
-// Copyright 2020 ETH Zurich and University of Bologna.
-// Licensed under the Apache License, Version 2.0, see LICENSE for details.
-// SPDX-License-Identifier: Apache-2.0
-
-// SW testbench for profiling GEMM kernels in different
-// floating point precisions (fp64, fp32, fp16), as well as
-// different memory layouts for matrices (transposed/not-transposed)
-
-#define BIST
-
-#include "dnn.h"
-#include "snrt.h"
-
-#include "data.h"
-
-// Padding of innermost dimension of a Matrix
-// Useful for preventing banking conflicts between cores
-// that are accessing different rows of the matrix
-#define MAT_ROW_PADDING 0
-
-// Padding in between matrices A, B for preventing
-// banking conflicts in the beginning
-#define MAT_PADDING 0
-
-void *share_ptr;
-
-int main() {
-    const gemm_layer_t l1_gemm_l = layer;
-
-    const uint32_t cluster_num = snrt_cluster_num();
-    const uint32_t cluster_id = snrt_cluster_idx();
-    const uint32_t compute_num = snrt_cluster_compute_core_num();
-    const uint32_t compute_id = snrt_global_core_idx();
-
-    void *mat_A, *mat_B, *mat_C;
-
-    uint32_t mat_A_size =
-        (l1_gemm_l.M * (l1_gemm_l.K + MAT_ROW_PADDING) + MAT_PADDING) *
-        l1_gemm_l.dtype;
-    uint32_t mat_B_size =
-        (l1_gemm_l.K + MAT_ROW_PADDING) * l1_gemm_l.N * l1_gemm_l.dtype;
-    uint32_t mat_C_size = l1_gemm_l.M * l1_gemm_l.N * l1_gemm_l.dtype;
-
-    uint32_t total_size = mat_A_size + mat_B_size + mat_C_size;
-
-    void *ptr;
-
-    if (compute_id == 0) {
-        ptr = snrt_l1alloc(total_size);
-        share_ptr = ptr;
-    }
-
-    snrt_cluster_hw_barrier();
-
-    ptr = share_ptr;
-
-    mat_A = ptr;
-    ptr += (l1_gemm_l.M * (l1_gemm_l.K + MAT_ROW_PADDING) + MAT_PADDING) *
-           l1_gemm_l.dtype;
-    mat_B = ptr;
-    ptr += (l1_gemm_l.K + MAT_ROW_PADDING) * l1_gemm_l.N * l1_gemm_l.dtype;
-    mat_C = ptr;
-    ptr += l1_gemm_l.M * l1_gemm_l.N * l1_gemm_l.dtype;
-
-    uint32_t errors = 0;
-
-    snrt_global_barrier();
-
-    if (snrt_is_dm_core()) {
-        snrt_dma_txid_t txid_A =
-            snrt_dma_start_2d(mat_A, l1_gemm_l.A, l1_gemm_l.dtype * l1_gemm_l.K,
-                              l1_gemm_l.dtype * (l1_gemm_l.K + MAT_ROW_PADDING),
-                              l1_gemm_l.dtype * l1_gemm_l.K, l1_gemm_l.M);
-        snrt_dma_txid_t txid_B =
-            snrt_dma_start_2d(mat_B, l1_gemm_l.B, l1_gemm_l.dtype * l1_gemm_l.K,
-                              l1_gemm_l.dtype * (l1_gemm_l.K + MAT_ROW_PADDING),
-                              l1_gemm_l.dtype * l1_gemm_l.K, l1_gemm_l.N);
-
-        snrt_dma_txid_t txid_C = snrt_dma_start_1d(
-            mat_C, l1_gemm_l.C, l1_gemm_l.dtype * l1_gemm_l.M * l1_gemm_l.N);
-
-        snrt_dma_wait_all();
-    }
-
-    snrt_cluster_hw_barrier();
-
-    if (snrt_is_compute_core()) {
-        const uint32_t setup_SSR = 1;
-
-        if (!l1_gemm_l.TA && !l1_gemm_l.TB) {
-            volatile uint32_t A_offset =
-                compute_id * (l1_gemm_l.K + MAT_ROW_PADDING) * l1_gemm_l.dtype;
-            volatile uint32_t C_offset =
-                compute_id * l1_gemm_l.N * l1_gemm_l.dtype;
-            volatile uint32_t ldA =
-                compute_num * (l1_gemm_l.K + MAT_ROW_PADDING);
-            volatile uint32_t ldB = l1_gemm_l.N + MAT_ROW_PADDING;
-            volatile uint32_t ldC = l1_gemm_l.N * compute_num;
-
-            snrt_mcycle();
-            gemm_fp64_opt(l1_gemm_l.M / compute_num, l1_gemm_l.N, l1_gemm_l.K,
-                          &mat_A[A_offset], ldA, l1_gemm_l.TA, mat_B, ldB,
-                          l1_gemm_l.TB, &mat_C[C_offset], ldC, &l1_gemm_l.ALPHA,
-                          setup_SSR);
-            snrt_mcycle();
-        } else if (!l1_gemm_l.TA && l1_gemm_l.TB) {
-            volatile uint32_t A_offset =
-                compute_id * (l1_gemm_l.K + MAT_ROW_PADDING) * l1_gemm_l.dtype;
-            volatile uint32_t C_offset =
-                compute_id * l1_gemm_l.N * l1_gemm_l.dtype;
-            volatile uint32_t ldA =
-                compute_num * (l1_gemm_l.K + MAT_ROW_PADDING);
-            volatile uint32_t ldB = l1_gemm_l.K + MAT_ROW_PADDING;
-            volatile uint32_t ldC = l1_gemm_l.N * compute_num;
-
-            snrt_mcycle();
-            switch (l1_gemm_l.dtype) {
-                case FP64:
-                    gemm_fp64_opt(l1_gemm_l.M / compute_num, l1_gemm_l.N,
-                                  l1_gemm_l.K, &mat_A[A_offset], ldA,
-                                  l1_gemm_l.TA, mat_B, ldB, l1_gemm_l.TB,
-                                  &mat_C[C_offset], ldC, &l1_gemm_l.ALPHA,
-                                  setup_SSR);
-                    break;
-                case FP32:
-                    gemm_fp32_opt(l1_gemm_l.M / compute_num, l1_gemm_l.N,
-                                  l1_gemm_l.K, &mat_A[A_offset], ldA, mat_B,
-                                  ldB, &mat_C[C_offset], ldC, &l1_gemm_l.ALPHA,
-                                  setup_SSR);
-                    break;
-                case FP16:
-                    if (l1_gemm_l.expand) {
-                        gemm_fp16_ex_opt(l1_gemm_l.M / compute_num, l1_gemm_l.N,
-                                         l1_gemm_l.K, &mat_A[A_offset], ldA,
-                                         mat_B, ldB, &mat_C[C_offset], ldC,
-                                         &l1_gemm_l.ALPHA, setup_SSR);
-                    } else {
-                        gemm_fp16_opt(l1_gemm_l.M / compute_num, l1_gemm_l.N,
-                                      l1_gemm_l.K, &mat_A[A_offset], ldA, mat_B,
-                                      ldB, &mat_C[C_offset], ldC,
-                                      &l1_gemm_l.ALPHA, setup_SSR);
-                    }
-                    break;
-                case FP8:
-                    gemm_fp8_ex_opt(l1_gemm_l.M / compute_num, l1_gemm_l.N,
-                                    l1_gemm_l.K, &mat_A[A_offset], ldA, mat_B,
-                                    ldB, &mat_C[C_offset], ldC,
-                                    &l1_gemm_l.ALPHA, setup_SSR);
-                    break;
-            }
-            snrt_mcycle();
-        } else if (l1_gemm_l.TA) {
-            printf("transpose TA not supported\n");
-        }
-        snrt_cluster_hw_barrier();
-    } else {
-        snrt_cluster_hw_barrier();
-    }
-    snrt_cluster_hw_barrier();
-
-#ifdef BIST
-
-    if (compute_id == 0) {
-        if (l1_gemm_l.dtype == FP64) {
-            for (uint32_t m = 0; m < l1_gemm_l.M; m++) {
-                double check = checksum[m];
-                double sum = 0.0;
-                for (uint32_t n = 0; n < l1_gemm_l.N; n++) {
-                    sum += ((double *)mat_C)[m * l1_gemm_l.N + n];
-                }
-                if (fabs(sum - check) > 0.001) {
-                    errors += l1_gemm_l.N;
-                }
-            }
-        } else if (l1_gemm_l.dtype == FP32) {
-            for (uint32_t m = 0; m < l1_gemm_l.M; m++) {
-                float check = checksum[m];
-                float sum = 0.0;
-                for (uint32_t n = 0; n < l1_gemm_l.N; n++) {
-                    sum += ((float *)mat_C)[m * l1_gemm_l.N + n];
-                }
-                if (fabs(sum - check) > 0.001) {
-                    errors += l1_gemm_l.N;
-                }
-            }
-        } else if (l1_gemm_l.dtype == FP16) {
-            for (uint32_t m = 0; m < l1_gemm_l.M; m++) {
-                __fp16 check = checksum[m];
-                float sum = 0.0;
-                for (uint32_t n = 0; n < l1_gemm_l.N; n++) {
-                    sum += ((__fp16 *)mat_C)[m * l1_gemm_l.N + n];
-                }
-                if (fabs(sum - check) > 0.05) {
-                    errors += l1_gemm_l.N;
-                }
-            }
-        } else if (l1_gemm_l.dtype == FP8) {
-            printf("No golden model yet for fp8!\n");
-        }
-        printf("%d/%d Errors\n", errors, l1_gemm_l.M * l1_gemm_l.N);
-    }
-
-#endif
-
-    // TODO: change back!!!
-    return 0;
-}
diff --git a/sw/dnn/src/dnn.h b/sw/dnn/src/dnn.h
index 3d0bea4682..e0d6e6e0e5 100644
--- a/sw/dnn/src/dnn.h
+++ b/sw/dnn/src/dnn.h
@@ -200,7 +200,6 @@ typedef struct network_single_cluster_t_ {
 #include "../flashattention_2/src/flashattention_2.h"
 #include "../fused_concat_linear/src/fused_concat_linear.h"
 #include "../gelu/src/gelu.h"
-#include "../gemm/src/gemm.h"
 #include "../layernorm/src/layernorm.h"
 #include "../maxpool/src/maxpool.h"
 #include "../softmax/src/softmax.h"
diff --git a/target/snitch_cluster/sw/apps/Makefile b/target/snitch_cluster/sw/apps/Makefile
index 3c4da406ab..a234b73906 100644
--- a/target/snitch_cluster/sw/apps/Makefile
+++ b/target/snitch_cluster/sw/apps/Makefile
@@ -13,7 +13,6 @@ SUBDIRS += dnn/batchnorm
 SUBDIRS += dnn/conv2d
 SUBDIRS += dnn/fusedconv
 SUBDIRS += dnn/gelu
-SUBDIRS += dnn/gemm
 SUBDIRS += dnn/layernorm
 SUBDIRS += dnn/maxpool
 SUBDIRS += dnn/softmax
diff --git a/target/snitch_cluster/sw/apps/dnn/gemm/Makefile b/target/snitch_cluster/sw/apps/dnn/gemm/Makefile
deleted file mode 100644
index 48a31215e5..0000000000
--- a/target/snitch_cluster/sw/apps/dnn/gemm/Makefile
+++ /dev/null
@@ -1,12 +0,0 @@
-# Copyright 2023 ETH Zurich and University of Bologna.
-# Licensed under the Apache License, Version 2.0, see LICENSE for details.
-# SPDX-License-Identifier: Apache-2.0
-#
-# Luca Colagrande <colluca@iis.ee.ethz.ch>
-
-APP ?= gemm
-
-include ../../../../../../sw/dnn/common.mk
-include ../../common.mk
-
-$(DEP): $(DATA_H)
\ No newline at end of file
diff --git a/target/snitch_cluster/sw/run.yaml b/target/snitch_cluster/sw/run.yaml
index cb0a386f5a..bf211effe0 100644
--- a/target/snitch_cluster/sw/run.yaml
+++ b/target/snitch_cluster/sw/run.yaml
@@ -75,7 +75,6 @@ runs:
     cmd: [../../../sw/blas/gemm/verify.py, "${sim_bin}", "${elf}"]
   - elf: apps/dnn/batchnorm/build/batchnorm.elf
   - elf: apps/dnn/maxpool/build/maxpool.elf
-  - elf: apps/dnn/gemm/build/gemm.elf
   # - elf: apps/dnn/conv2d/build/conv2d.elf # Fails with wrong results
   #   cmd: [../../../sw/dnn/conv2d/verify.py, "${sim_bin}", "${elf}"]
   # - elf: apps/dnn/fusedconv/build/fusedconv.elf # Fails with wrong results