diff --git a/sw/dnn/decoder/data/gpt-3xl/gpt-3xl-fp16.json b/sw/dnn/decoder/data/gpt-3xl/gpt-3xl-fp16.json
new file mode 100644
index 000000000..dcc871106
--- /dev/null
+++ b/sw/dnn/decoder/data/gpt-3xl/gpt-3xl-fp16.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 40,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 2048
+            },
+            eps: 1e-5,
+            prec: "FP16",
+            n_tiles: 66,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 2,
+            k_tiles: 8,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 128,
+            K: 2048,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        },
+        flashattention_2: {
+            L: 2048,
+            S: 2048,
+            d: 128,
+            B_r: 16,
+            B_c: 16,
+            dtype: "FP16",
+            use_mask: true,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 128],
+            output_shape: [16, 128],
+            dtype: "FP16",
+            gemm_implementation: "gemm_fp16_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 2048
+            },
+            eps: 1e-5,
+            prec: "FP16",
+            n_tiles: 66,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256, // number of tiles in M dimension
+            n_tiles: 2, // number of tiles in N dimension
+            k_tiles: 8, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 2048,
+            N: 128,
+            K: 2048,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 32,
+            k_tiles: 32,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 2048,
+            K: 8192,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/decoder/data/gpt-3xl/gpt-3xl-fp32.json b/sw/dnn/decoder/data/gpt-3xl/gpt-3xl-fp32.json
new file mode 100644
index 000000000..8ff3e596f
--- /dev/null
+++ b/sw/dnn/decoder/data/gpt-3xl/gpt-3xl-fp32.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 40,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 2048
+            },
+            eps: 1e-5,
+            prec: "FP32",
+            n_tiles: 132,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 2,
+            k_tiles: 16,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 128,
+            K: 2048,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        },
+        flashattention_2: {
+            L: 2048,
+            S: 2048,
+            d: 128,
+            B_r: 8,
+            B_c: 8,
+            dtype: "FP32",
+            use_mask: true,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 128],
+            output_shape: [16, 128],
+            dtype: "FP32",
+            gemm_implementation: "gemm_fp32_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 2048
+            },
+            eps: 1e-5,
+            prec: "FP32",
+            n_tiles: 132,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256, // number of tiles in M dimension
+            n_tiles: 2, // number of tiles in N dimension
+            k_tiles: 16, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 2048,
+            N: 128,
+            K: 2048,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 32,
+            k_tiles: 64,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 2048,
+            K: 8192,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/decoder/data/gpt-3xl/gpt-3xl-fp8.json b/sw/dnn/decoder/data/gpt-3xl/gpt-3xl-fp8.json
new file mode 100644
index 000000000..1627f6512
--- /dev/null
+++ b/sw/dnn/decoder/data/gpt-3xl/gpt-3xl-fp8.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 40,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 2048
+            },
+            eps: 1e-5,
+            prec: "FP8",
+            n_tiles: 33,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 2,
+            k_tiles: 4,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 128,
+            K: 2048,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        },
+        flashattention_2: {
+            L: 2048,
+            S: 2048,
+            d: 128,
+            B_r: 32,
+            B_c: 32,
+            dtype: "FP8",
+            use_mask: true,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 128],
+            output_shape: [16, 128],
+            dtype: "FP8",
+            gemm_implementation: "gemm_fp8_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 2048
+            },
+            eps: 1e-5,
+            prec: "FP8",
+            n_tiles: 33,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256, // number of tiles in M dimension
+            n_tiles: 2, // number of tiles in N dimension
+            k_tiles: 4, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 2048,
+            N: 128,
+            K: 2048,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 32,
+            k_tiles: 16,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 2048,
+            K: 8192,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/decoder/data/gpt-j/gpt-j-fp16.json b/sw/dnn/decoder/data/gpt-j/gpt-j-fp16.json
new file mode 100644
index 000000000..8d63923ef
--- /dev/null
+++ b/sw/dnn/decoder/data/gpt-j/gpt-j-fp16.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 28,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 4096
+            },
+            eps: 1e-5,
+            prec: "FP16",
+            n_tiles: 132,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 4,
+            k_tiles: 16,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 256,
+            K: 4096,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        },
+        flashattention_2: {
+            L: 2048,
+            S: 2048,
+            d: 256,
+            B_r: 8,
+            B_c: 8,
+            dtype: "FP16",
+            use_mask: true,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 256],
+            output_shape: [16, 128],
+            dtype: "FP16",
+            gemm_implementation: "gemm_fp16_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 4096
+            },
+            eps: 1e-5,
+            prec: "FP16",
+            n_tiles: 132,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256, // number of tiles in M dimension
+            n_tiles: 4, // number of tiles in N dimension
+            k_tiles: 16, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 2048,
+            N: 256,
+            K: 4096,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 64,
+            k_tiles: 64,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 4096,
+            K: 16384,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/decoder/data/gpt-j/gpt-j-fp32.json b/sw/dnn/decoder/data/gpt-j/gpt-j-fp32.json
new file mode 100644
index 000000000..e43fd2f33
--- /dev/null
+++ b/sw/dnn/decoder/data/gpt-j/gpt-j-fp32.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 28,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 4096
+            },
+            eps: 1e-5,
+            prec: "FP32",
+            n_tiles: 264,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 4,
+            k_tiles: 32,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 256,
+            K: 4096,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        },
+        flashattention_2: {
+            L: 2048,
+            S: 2048,
+            d: 256,
+            B_r: 4,
+            B_c: 4,
+            dtype: "FP32",
+            use_mask: true,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 256],
+            output_shape: [16, 128],
+            dtype: "FP32",
+            gemm_implementation: "gemm_fp32_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 4096
+            },
+            eps: 1e-5,
+            prec: "FP32",
+            n_tiles: 264,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256, // number of tiles in M dimension
+            n_tiles: 4, // number of tiles in N dimension
+            k_tiles: 32, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 2048,
+            N: 256,
+            K: 4096,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 64,
+            k_tiles: 128,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 4096,
+            K: 16384,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/decoder/data/gpt-j/gpt-j-fp8.json b/sw/dnn/decoder/data/gpt-j/gpt-j-fp8.json
new file mode 100644
index 000000000..fa8948ddf
--- /dev/null
+++ b/sw/dnn/decoder/data/gpt-j/gpt-j-fp8.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 28,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 4096
+            },
+            eps: 1e-5,
+            prec: "FP8",
+            n_tiles: 66,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 4,
+            k_tiles: 8,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 256,
+            K: 4096,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        },
+        flashattention_2: {
+            L: 2048,
+            S: 2048,
+            d: 256,
+            B_r: 16,
+            B_c: 16,
+            dtype: "FP8",
+            use_mask: true,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 256],
+            output_shape: [16, 128],
+            dtype: "FP8",
+            gemm_implementation: "gemm_fp8_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 2048,
+                embeddings: 4096
+            },
+            eps: 1e-5,
+            prec: "FP8",
+            n_tiles: 66,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256, // number of tiles in M dimension
+            n_tiles: 4, // number of tiles in N dimension
+            k_tiles: 8, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 2048,
+            N: 256,
+            K: 4096,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 256,
+            n_tiles: 64,
+            k_tiles: 32,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 2048,
+            N: 4096,
+            K: 16384,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/encoder/data/vit-b/vit-b-fp16.json b/sw/dnn/encoder/data/vit-b/vit-b-fp16.json
new file mode 100644
index 000000000..803c736d4
--- /dev/null
+++ b/sw/dnn/encoder/data/vit-b/vit-b-fp16.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 12,
+    mha: {
+        num_heads: 12,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 768
+            },
+            eps: 1e-5,
+            prec: "FP16",
+            n_tiles: 3,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 1,
+            k_tiles: 3,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 64,
+            K: 768,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        },
+        flashattention_2: {
+            L: 197,
+            S: 197,
+            d: 64,
+            B_r: 32,
+            B_c: 32,
+            dtype: "FP16",
+            use_mask: false,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 12,
+            input_shape: [16, 64],
+            output_shape: [16, 128],
+            dtype: "FP16",
+            gemm_implementation: "gemm_fp16_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 768
+            },
+            eps: 1e-5,
+            prec: "FP16",
+            n_tiles: 3,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25, // number of tiles in M dimension
+            n_tiles: 1, // number of tiles in N dimension
+            k_tiles: 3, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 197,
+            N: 64,
+            K: 768,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 12,
+            k_tiles: 12,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 768,
+            K: 3072,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/encoder/data/vit-b/vit-b-fp32.json b/sw/dnn/encoder/data/vit-b/vit-b-fp32.json
new file mode 100644
index 000000000..61a363da1
--- /dev/null
+++ b/sw/dnn/encoder/data/vit-b/vit-b-fp32.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 12,
+    mha: {
+        num_heads: 12,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 768
+            },
+            eps: 1e-5,
+            prec: "FP32",
+            n_tiles: 5,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 1,
+            k_tiles: 6,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 64,
+            K: 768,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        },
+        flashattention_2: {
+            L: 197,
+            S: 197,
+            d: 64,
+            B_r: 16,
+            B_c: 16,
+            dtype: "FP32",
+            use_mask: false,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 12,
+            input_shape: [16, 64],
+            output_shape: [16, 128],
+            dtype: "FP32",
+            gemm_implementation: "gemm_fp32_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 768
+            },
+            eps: 1e-5,
+            prec: "FP32",
+            n_tiles: 5,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25, // number of tiles in M dimension
+            n_tiles: 1, // number of tiles in N dimension
+            k_tiles: 6, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 197,
+            N: 64,
+            K: 768,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 12,
+            k_tiles: 24,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 768,
+            K: 3072,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/encoder/data/vit-b/vit-b-fp8.json b/sw/dnn/encoder/data/vit-b/vit-b-fp8.json
new file mode 100644
index 000000000..c4953d4e2
--- /dev/null
+++ b/sw/dnn/encoder/data/vit-b/vit-b-fp8.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 12,
+    mha: {
+        num_heads: 12,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 768
+            },
+            eps: 1e-5,
+            prec: "FP8",
+            n_tiles: 2,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 1,
+            k_tiles: 2,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 64,
+            K: 768,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        },
+        flashattention_2: {
+            L: 197,
+            S: 197,
+            d: 64,
+            B_r: 64,
+            B_c: 64,
+            dtype: "FP8",
+            use_mask: false,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 12,
+            input_shape: [16, 64],
+            output_shape: [16, 128],
+            dtype: "FP8",
+            gemm_implementation: "gemm_fp8_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 768
+            },
+            eps: 1e-5,
+            prec: "FP8",
+            n_tiles: 2,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25, // number of tiles in M dimension
+            n_tiles: 1, // number of tiles in N dimension
+            k_tiles: 2, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 197,
+            N: 64,
+            K: 768,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 12,
+            k_tiles: 6,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 768,
+            K: 3072,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/encoder/data/vit-h/vit-h-fp16.json b/sw/dnn/encoder/data/vit-h/vit-h-fp16.json
new file mode 100644
index 000000000..60e84585b
--- /dev/null
+++ b/sw/dnn/encoder/data/vit-h/vit-h-fp16.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 32,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1280
+            },
+            eps: 1e-5,
+            prec: "FP16",
+            n_tiles: 8,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 1,
+            k_tiles: 10,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 64,
+            K: 1280,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        },
+        flashattention_2: {
+            L: 197,
+            S: 197,
+            d: 80,
+            B_r: 32,
+            B_c: 32,
+            dtype: "FP16",
+            use_mask: false,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 80],
+            output_shape: [16, 128],
+            dtype: "FP16",
+            gemm_implementation: "gemm_fp16_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1280
+            },
+            eps: 1e-5,
+            prec: "FP16",
+            n_tiles: 8,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25, // number of tiles in M dimension
+            n_tiles: 1, // number of tiles in N dimension
+            k_tiles: 8, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 197,
+            N: 64,
+            K: 1280,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 30,
+            k_tiles: 20,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 1280,
+            K: 5120,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/encoder/data/vit-h/vit-h-fp32.json b/sw/dnn/encoder/data/vit-h/vit-h-fp32.json
new file mode 100644
index 000000000..ada7da7fc
--- /dev/null
+++ b/sw/dnn/encoder/data/vit-h/vit-h-fp32.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 32,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1280
+            },
+            eps: 1e-5,
+            prec: "FP32",
+            n_tiles: 15,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 1,
+            k_tiles: 20,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 64,
+            K: 1280,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        },
+        flashattention_2: {
+            L: 197,
+            S: 197,
+            d: 80,
+            B_r: 16,
+            B_c: 16,
+            dtype: "FP32",
+            use_mask: false,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 80],
+            output_shape: [16, 128],
+            dtype: "FP32",
+            gemm_implementation: "gemm_fp32_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1280
+            },
+            eps: 1e-5,
+            prec: "FP32",
+            n_tiles: 15,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25, // number of tiles in M dimension
+            n_tiles: 1, // number of tiles in N dimension
+            k_tiles: 15, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 197,
+            N: 64,
+            K: 1280,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 30,
+            k_tiles: 40,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 1280,
+            K: 5120,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/encoder/data/vit-h/vit-h-fp8.json b/sw/dnn/encoder/data/vit-h/vit-h-fp8.json
new file mode 100644
index 000000000..ed86d3cae
--- /dev/null
+++ b/sw/dnn/encoder/data/vit-h/vit-h-fp8.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 32,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1280
+            },
+            eps: 1e-5,
+            prec: "FP8",
+            n_tiles: 4,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 1,
+            k_tiles: 5,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 64,
+            K: 1280,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        },
+        flashattention_2: {
+            L: 197,
+            S: 197,
+            d: 80,
+            B_r: 64,
+            B_c: 64,
+            dtype: "FP8",
+            use_mask: false,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 80],
+            output_shape: [16, 128],
+            dtype: "FP8",
+            gemm_implementation: "gemm_fp8_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1280
+            },
+            eps: 1e-5,
+            prec: "FP8",
+            n_tiles: 4,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25, // number of tiles in M dimension
+            n_tiles: 1, // number of tiles in N dimension
+            k_tiles: 4, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 197,
+            N: 64,
+            K: 1280,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 30,
+            k_tiles: 10,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 1280,
+            K: 5120,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/encoder/data/vit-l/vit-l-fp16.json b/sw/dnn/encoder/data/vit-l/vit-l-fp16.json
new file mode 100644
index 000000000..d53a57b28
--- /dev/null
+++ b/sw/dnn/encoder/data/vit-l/vit-l-fp16.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 24,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1024
+            },
+            eps: 1e-5,
+            prec: "FP16",
+            n_tiles: 6,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 1,
+            k_tiles: 8,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 64,
+            K: 1024,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        },
+        flashattention_2: {
+            L: 197,
+            S: 197,
+            d: 64,
+            B_r: 32,
+            B_c: 32,
+            dtype: "FP16",
+            use_mask: false,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 64],
+            output_shape: [16, 128],
+            dtype: "FP16",
+            gemm_implementation: "gemm_fp16_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1024
+            },
+            eps: 1e-5,
+            prec: "FP16",
+            n_tiles: 6,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25, // number of tiles in M dimension
+            n_tiles: 1, // number of tiles in N dimension
+            k_tiles: 6, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 197,
+            N: 64,
+            K: 1024,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 24,
+            k_tiles: 16,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 1024,
+            K: 4096,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp16_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/encoder/data/vit-l/vit-l-fp32.json b/sw/dnn/encoder/data/vit-l/vit-l-fp32.json
new file mode 100644
index 000000000..5c9d4dcb6
--- /dev/null
+++ b/sw/dnn/encoder/data/vit-l/vit-l-fp32.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 24,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1024
+            },
+            eps: 1e-5,
+            prec: "FP32",
+            n_tiles: 12,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 1,
+            k_tiles: 16,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 64,
+            K: 1024,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        },
+        flashattention_2: {
+            L: 197,
+            S: 197,
+            d: 64,
+            B_r: 16,
+            B_c: 16,
+            dtype: "FP32",
+            use_mask: false,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 64],
+            output_shape: [16, 128],
+            dtype: "FP32",
+            gemm_implementation: "gemm_fp32_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1024
+            },
+            eps: 1e-5,
+            prec: "FP32",
+            n_tiles: 12,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25, // number of tiles in M dimension
+            n_tiles: 1, // number of tiles in N dimension
+            k_tiles: 12, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 197,
+            N: 64,
+            K: 1024,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 24,
+            k_tiles: 32,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 1024,
+            K: 4096,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp32_opt"
+        }
+    }
+}
\ No newline at end of file
diff --git a/sw/dnn/encoder/data/vit-l/vit-l-fp8.json b/sw/dnn/encoder/data/vit-l/vit-l-fp8.json
new file mode 100644
index 000000000..58a1bece1
--- /dev/null
+++ b/sw/dnn/encoder/data/vit-l/vit-l-fp8.json
@@ -0,0 +1,108 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    num_layers : 24,
+    mha: {
+        num_heads: 16,
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1024
+            },
+            eps: 1e-5,
+            prec: "FP8",
+            n_tiles: 3,
+            implementation: "OPT"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 1,
+            k_tiles: 4,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 64,
+            K: 1024,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        },
+        flashattention_2: {
+            L: 197,
+            S: 197,
+            d: 64,
+            B_r: 64,
+            B_c: 64,
+            dtype: "FP8",
+            use_mask: false,
+            baseline: false
+        },
+        'fused_concat_linear': {
+            num_inputs: 16,
+            input_shape: [16, 64],
+            output_shape: [16, 128],
+            dtype: "FP8",
+            gemm_implementation: "gemm_fp8_opt"
+        }
+    },
+    mlp: {
+        layernorm: {
+            input_dim: {
+                batch_size: 1,
+                seq_len: 197,
+                embeddings: 1024
+            },
+            eps: 1e-5,
+            prec: "FP8",
+            n_tiles: 3,
+            implementation: "OPT"
+        },
+        'fused_linear_gelu': {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25, // number of tiles in M dimension
+            n_tiles: 1, // number of tiles in N dimension
+            k_tiles: 3, // number of tiles in K dimension
+            load_a: 1,
+            load_b: 1,
+            load_c: 1,
+            transa: false,
+            transb: true, // must be true for SIMD
+            M: 197,
+            N: 64,
+            K: 1024,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        },
+        gemm: {
+            setup_ssr: 1,
+            parallelize_m: 0,
+            parallelize_k: 0,
+            m_tiles: 25,
+            n_tiles: 24,
+            k_tiles: 8,
+            load_a: 1,
+            load_b: 1,
+            load_c: 0,
+            transa: false,
+            transb: true,
+            M: 197,
+            N: 1024,
+            K: 4096,
+            alpha: 1,
+            beta: 0,
+            gemm_fp: "gemm_fp8_opt"
+        }
+    }
+}
\ No newline at end of file