diff --git a/TracyLua.hpp b/TracyLua.hpp
index 5490467c..e4e0e71c 100644
--- a/TracyLua.hpp
+++ b/TracyLua.hpp
@@ -150,9 +150,9 @@ static tracy_force_inline void SendLuaCallstack( lua_State* L, uint32_t depth )
     const char* func[64];
     uint32_t fsz[64];
     uint32_t ssz[64];
-    uint32_t spaceNeeded = 4;     // cnt
+    uint16_t spaceNeeded = 2;     // cnt
 
-    uint32_t cnt;
+    uint8_t cnt;
     for( cnt=0; cnt<depth; cnt++ )
     {
         if( lua_getstack( L, cnt+1, dbg+cnt ) == 0 ) break;
@@ -162,22 +162,24 @@ static tracy_force_inline void SendLuaCallstack( lua_State* L, uint32_t depth )
         ssz[cnt] = uint32_t( strlen( dbg[cnt].source ) );
         spaceNeeded += fsz[cnt] + ssz[cnt];
     }
-    spaceNeeded += cnt * ( 4 + 4 + 4 );     // source line, function string length, source string length
+    spaceNeeded += cnt * ( 4 + 2 + 2 );     // source line, function string length, source string length
 
-    auto ptr = (char*)tracy_malloc( spaceNeeded + 4 );
+    auto ptr = (char*)tracy_malloc( spaceNeeded + 2 );
     auto dst = ptr;
-    memcpy( dst, &spaceNeeded, 4 ); dst += 4;
-    memcpy( dst, &cnt, 4 ); dst += 4;
-    for( uint32_t i=0; i<cnt; i++ )
+    memcpy( dst, &spaceNeeded, 2 ); dst += 2;
+    memcpy( dst, &cnt, 1 ); dst++;
+    for( uint8_t i=0; i<cnt; i++ )
     {
         const uint32_t line = dbg[i].currentline;
         memcpy( dst, &line, 4 ); dst += 4;
-        memcpy( dst, fsz+i, 4 ); dst += 4;
+        assert( fsz[i] <= std::numeric_limits<uint16_t>::max() );
+        memcpy( dst, fsz+i, 2 ); dst += 2;
         memcpy( dst, func[i], fsz[i] ); dst += fsz[i];
-        memcpy( dst, ssz+i, 4 ); dst += 4;
+        assert( ssz[i] <= std::numeric_limits<uint16_t>::max() );
+        memcpy( dst, ssz+i, 2 ); dst += 2;
         memcpy( dst, dbg[i].source, ssz[i] ), dst += ssz[i];
     }
-    assert( dst - ptr == spaceNeeded + 4 );
+    assert( dst - ptr == spaceNeeded + 2 );
 
     TracyLfqPrepare( QueueType::CallstackAlloc );
     MemWrite( &item->callstackAlloc.ptr, (uint64_t)ptr );
diff --git a/client/TracyProfiler.cpp b/client/TracyProfiler.cpp
index 0aa1326a..8d2da1a8 100644
--- a/client/TracyProfiler.cpp
+++ b/client/TracyProfiler.cpp
@@ -2307,15 +2307,15 @@ void Profiler::SendCallstackAlloc( uint64_t _ptr )
     MemWrite( &item.hdr.type, QueueType::CallstackAllocPayload );
     MemWrite( &item.stringTransfer.ptr, _ptr );
 
-    const auto len = *((uint32_t*)ptr);
-    assert( len <= std::numeric_limits<uint16_t>::max() );
-    const auto l16 = uint16_t( len );
+    uint16_t len;
+    memcpy( &len, ptr, 2 );
+    ptr += 2;
 
-    NeedDataSize( QueueDataSize[(int)QueueType::CallstackAllocPayload] + sizeof( l16 ) + l16 );
+    NeedDataSize( QueueDataSize[(int)QueueType::CallstackAllocPayload] + sizeof( len ) + len );
 
     AppendDataUnsafe( &item, QueueDataSize[(int)QueueType::CallstackAllocPayload] );
-    AppendDataUnsafe( &l16, sizeof( l16 ) );
-    AppendDataUnsafe( ptr + 4, l16 );
+    AppendDataUnsafe( &len, sizeof( len ) );
+    AppendDataUnsafe( ptr, len );
 }
 
 void Profiler::SendCallstackFrame( uint64_t ptr )
diff --git a/common/TracyProtocol.hpp b/common/TracyProtocol.hpp
index 2160811e..634b5846 100644
--- a/common/TracyProtocol.hpp
+++ b/common/TracyProtocol.hpp
@@ -9,7 +9,7 @@ namespace tracy
 
 constexpr unsigned Lz4CompressBound( unsigned isize ) { return isize + ( isize / 255 ) + 16; }
 
-enum : uint32_t { ProtocolVersion = 35 };
+enum : uint32_t { ProtocolVersion = 36 };
 enum : uint32_t { BroadcastVersion = 1 };
 
 using lz4sz_t = uint32_t;
diff --git a/server/TracyWorker.cpp b/server/TracyWorker.cpp
index a21b275a..5d19a8b6 100644
--- a/server/TracyWorker.cpp
+++ b/server/TracyWorker.cpp
@@ -3487,16 +3487,17 @@ void Worker::AddCallstackPayload( uint64_t ptr, const char* _data, size_t _sz )
 void Worker::AddCallstackAllocPayload( uint64_t ptr, const char* data, size_t _sz )
 {
     CallstackFrameId stack[64];
-    const auto sz = *(uint32_t*)data; data += 4;
+    uint8_t sz;
+    memcpy( &sz, data, 1 ); data++;
     assert( sz <= 64 );
-    for( uint32_t i=0; i<sz; i++ )
+    for( uint8_t i=0; i<sz; i++ )
     {
-        uint32_t sz;
+        uint16_t sz;
         CallstackFrame cf;
         memcpy( &cf.line, data, 4 ); data += 4;
-        memcpy( &sz, data, 4 ); data += 4;
+        memcpy( &sz, data, 2 ); data += 2;
         cf.name = StoreString( data, sz ).idx; data += sz;
-        memcpy( &sz, data, 4 ); data += 4;
+        memcpy( &sz, data, 2 ); data += 2;
         cf.file = StoreString( data, sz ).idx; data += sz;
         cf.symAddr = 0;
         CallstackFrameData cfd = { &cf, 1 };