diff --git a/src/gallium/drivers/ethosu/ethosu_cmd.c b/src/gallium/drivers/ethosu/ethosu_cmd.c
index 6cf1eaddde2..4eeefc557f7 100644
--- a/src/gallium/drivers/ethosu/ethosu_cmd.c
+++ b/src/gallium/drivers/ethosu/ethosu_cmd.c
@@ -721,6 +721,23 @@ simplified_elementwise_add_sub_scale(
    *out_out_scale = ethosu_quantize_scale(output_rescale_val, out_out_shift, false);
 }
 
+static void
+elementwise_mul_scale(
+   struct ethosu_subgraph *subgraph,
+   double input1_scale,
+   double input2_scale,
+   double output_scale)
+{
+   double output_rescale;
+   int32_t ofm_scale, ofm_shift;
+
+   output_rescale = (input1_scale * input2_scale) / output_scale;
+   ofm_scale = ethosu_quantize_scale(output_rescale, &ofm_shift, false);
+
+   /* OFM_SCALE: output scale with shift */
+   EMIT1(NPU_SET_OFM_SCALE, ofm_shift, ofm_scale);
+}
+
 /*
  * U85 uses "simplified" mode (from Vela simplified_elementwise_add_sub_scale):
  *   Both operands are independently rescaled.  OPA_SCALE and OPB_SCALE each
@@ -765,27 +782,39 @@ static void
 emit_eltwise(struct ethosu_subgraph *subgraph, struct ethosu_operation *operation)
 {
    bool has_scalar = operation->ifm2.scalar != 0;
-   enum ethosu_op_to_scale op_to_scale;
+   enum ethosu_op_to_scale op_to_scale = OP_NONE;
 
-   if (ethosu_ml_device(subgraph->base.device)->is_u65) {
-      op_to_scale = eltwise_emit_ofm_scaling(
-         subgraph,
-         operation->ifm.scale,
-         operation->ifm2.scale,
-         operation->ofm.scale);
-   } else {
-      op_to_scale = eltwise_emit_ofm_scaling_u85(
-         subgraph,
-         operation->ifm.scale,
-         operation->ifm2.scale,
-         operation->ofm.scale);
-   }
+   switch (operation->eltwise.type) {
+   case ETHOSU_ELTWISE_TYPE_MUL:
+      elementwise_mul_scale(subgraph, operation->ifm.scale,
+                            operation->ifm2.scale,
+                            operation->ofm.scale);
+      break;
+   case ETHOSU_ELTWISE_TYPE_ADD:
+      if (ethosu_ml_device(subgraph->base.device)->is_u65) {
+         op_to_scale = eltwise_emit_ofm_scaling(
+            subgraph,
+            operation->ifm.scale,
+            operation->ifm2.scale,
+            operation->ofm.scale);
+      } else {
+         op_to_scale = eltwise_emit_ofm_scaling_u85(
+            subgraph,
+            operation->ifm.scale,
+            operation->ifm2.scale,
+            operation->ofm.scale);
+      }
 
-   if (operation->eltwise.ifm_reversed) {
-      if (op_to_scale == OP_A)
-         op_to_scale = OP_B;
-      else
-         op_to_scale = OP_A;
+      if (operation->eltwise.ifm_reversed) {
+         if (op_to_scale == OP_A)
+            op_to_scale = OP_B;
+         else
+            op_to_scale = OP_A;
+      }
+      break;
+   default:
+      assert(0);
+      break;
    }
 
    emit_common(subgraph, operation, op_to_scale);
@@ -832,7 +861,7 @@ emit_operation_code(struct ethosu_subgraph *subgraph, struct ethosu_operation *o
       EMIT0(NPU_OP_POOL, operation->pooling.type);
       break;
    case ETHOSU_OPERATION_TYPE_ELTWISE:
-      EMIT0(NPU_OP_ELEMENTWISE, 0x1);
+      EMIT0(NPU_OP_ELEMENTWISE, operation->eltwise.type);
       break;
    case ETHOSU_OPERATION_TYPE_DMA:
       EMIT0(NPU_OP_DMA_START, 0x0);
diff --git a/src/gallium/drivers/ethosu/ethosu_lower.c b/src/gallium/drivers/ethosu/ethosu_lower.c
index 16c793a1564..49818a5f0a6 100644
--- a/src/gallium/drivers/ethosu/ethosu_lower.c
+++ b/src/gallium/drivers/ethosu/ethosu_lower.c
@@ -519,6 +519,14 @@ ethosu_lower_graph(struct ethosu_subgraph *subgraph,
 
       case PIPE_ML_OPERATION_TYPE_ADD: {
          ethosu_lower_eltwise(subgraph, &poperations[i], &operation);
+         operation.eltwise.type = ETHOSU_ELTWISE_TYPE_ADD;
+         util_dynarray_append(&subgraph->operations, operation);
+         break;
+      }
+
+      case PIPE_ML_OPERATION_TYPE_MUL: {
+         ethosu_lower_eltwise(subgraph, &poperations[i], &operation);
+         operation.eltwise.type = ETHOSU_ELTWISE_TYPE_MUL;
          util_dynarray_append(&subgraph->operations, operation);
          break;
       }
diff --git a/src/gallium/drivers/ethosu/ethosu_ml.c b/src/gallium/drivers/ethosu/ethosu_ml.c
index d8f9f7e23c2..aa810c12842 100644
--- a/src/gallium/drivers/ethosu/ethosu_ml.c
+++ b/src/gallium/drivers/ethosu/ethosu_ml.c
@@ -140,6 +140,7 @@ ethosu_ml_operation_supported(struct pipe_ml_device *pdevice,
 
       break;
    }
+   case PIPE_ML_OPERATION_TYPE_MUL:
    case PIPE_ML_OPERATION_TYPE_ADD:
    case PIPE_ML_OPERATION_TYPE_POOLING:
    case PIPE_ML_OPERATION_TYPE_STRIDED_SLICE:
diff --git a/src/gallium/drivers/ethosu/ethosu_ml.h b/src/gallium/drivers/ethosu/ethosu_ml.h
index 4d57d2c924a..4530997fa74 100644
--- a/src/gallium/drivers/ethosu/ethosu_ml.h
+++ b/src/gallium/drivers/ethosu/ethosu_ml.h
@@ -42,6 +42,31 @@ enum ethosu_operation_type {
    ETHOSU_OPERATION_TYPE_DMA,
 };
 
+enum ethosu_eltwise_type {
+   ETHOSU_ELTWISE_TYPE_MUL,
+   ETHOSU_ELTWISE_TYPE_ADD,
+   ETHOSU_ELTWISE_TYPE_SUB,
+   ETHOSU_ELTWISE_TYPE_MIN,
+   ETHOSU_ELTWISE_TYPE_MAX,
+   ETHOSU_ELTWISE_TYPE_LRELU,
+   ETHOSU_ELTWISE_TYPE_ABS,
+   /* U85 only */
+   ETHOSU_ELTWISE_TYPE_CLZ,
+   ETHOSU_ELTWISE_TYPE_SHR,
+   ETHOSU_ELTWISE_TYPE_SHL,
+   ETHOSU_ELTWISE_TYPE_LSR,
+   ETHOSU_ELTWISE_TYPE_DIV,
+   ETHOSU_ELTWISE_TYPE_CMP_EQ = 0x10,
+   ETHOSU_ELTWISE_TYPE_CMP_NE,
+   ETHOSU_ELTWISE_TYPE_CMP_GE,
+   ETHOSU_ELTWISE_TYPE_CMP_GT,
+   ETHOSU_ELTWISE_TYPE_CMP_AND = 0x21,
+   ETHOSU_ELTWISE_TYPE_CMP_OR,
+   ETHOSU_ELTWISE_TYPE_CMP_XOR,
+   ETHOSU_ELTWISE_TYPE_CMP_NOT,
+   ETHOSU_ELTWISE_TYPE_CMP_AND_NOT = 0x2A,
+};
+
 struct ethosu_tile_box {
    unsigned height_0;     /* The height of tile 0 */
    unsigned height_1;     /* The height of tile 1, 0 if unused */
@@ -173,6 +198,7 @@ struct ethosu_operation {
       } pooling;
 
       struct {
+         enum ethosu_eltwise_type type;
          uint16_t activation_min;
          unsigned lut_bytes;
          bool ifm_reversed;