diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
index 04ec5117f..57a0981b8 100644
--- a/ggml/src/ggml-metal.m
+++ b/ggml/src/ggml-metal.m
@@ -1973,7 +1973,7 @@ static void ggml_metal_encode_node(
                             [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
                             [encoder setBytes:&r2      length:sizeof(r2)   atIndex:15];
                             [encoder setBytes:&r3      length:sizeof(r3)   atIndex:16];
-                            [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                            [encoder setThreadgroupMemoryLength:(4096 + 2048) atIndex:0];
                             [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                         } else {
                             int nth0 = 32;
diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal.metal
index 413661c8a..e0eb37c5c 100644
--- a/ggml/src/ggml-metal.metal
+++ b/ggml/src/ggml-metal.metal
@@ -6310,8 +6310,8 @@ kernel void kernel_mul_mm(device const  uchar * src0,
                           uint                  tiitg[[thread_index_in_threadgroup]],
                           uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    threadgroup T     * sa = (threadgroup T     *)(shared_memory);
-    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
+    threadgroup T    * sa = (threadgroup T    *)(shared_memory);
+    threadgroup half * sb = (threadgroup half *)(shared_memory + 4096);
 
     const uint r0 = tgpig.y;
     const uint r1 = tgpig.x;
@@ -6325,12 +6325,12 @@ kernel void kernel_mul_mm(device const  uchar * src0,
     short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
     short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
 
-    simdgroup_T8x8     ma[4];
-    simdgroup_float8x8 mb[2];
-    simdgroup_float8x8 mc[8];
+    simdgroup_T8x8    ma[4];
+    simdgroup_half8x8 mb[2];
+    simdgroup_half8x8 mc[8];
 
     for (short i = 0; i < 8; i++){
-        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+        mc[i] = make_filled_simdgroup_matrix<half, 8>(0.h);
     }
 
     short il = (tiitg % THREAD_PER_ROW);
@@ -6361,17 +6361,17 @@ kernel void kernel_mul_mm(device const  uchar * src0,
             +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
         }
 
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL)*8*32 + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y);
+        *(threadgroup half2x4 *)(sb + (tiitg%THREAD_PER_COL)*8*32 + 8*(tiitg/THREAD_PER_COL)) = (half2x4)(*((device float2x4 *)y));
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
-        x  = (il < 2) ? x + (2+nl-1)/nl : x;
+        x  = (il <  2) ? x + (2 + nl - 1)/nl : x;
         y += BLOCK_SIZE_K;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup T     * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
-        threadgroup float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
+        threadgroup T    * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
+        threadgroup half * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
 
         #pragma unroll(4)
         for (short ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
@@ -6399,15 +6399,22 @@ kernel void kernel_mul_mm(device const  uchar * src0,
         device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) \
                                + (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
         for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
+            // cast to f32
+            simdgroup_float8x8 mc_f32(1.0f);
+            simdgroup_multiply(mc_f32, mc[i], mc_f32);
+            simdgroup_store(mc_f32, C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
+            //simdgroup_store(mc[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
         }
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
         threadgroup float * temp_str = ((threadgroup float *) shared_memory) \
-                                      + 32 * (sgitg&1) + (16 * (sgitg>>1))*BLOCK_SIZE_M;
+                                       + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
         for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
+            simdgroup_float8x8 mc_f32(1.0f);
+            simdgroup_multiply(mc_f32, mc[i], mc_f32);
+            simdgroup_store(mc_f32, temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
+            //simdgroup_store(mc[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
         }
 
         threadgroup_barrier(mem_flags::mem_threadgroup);