diff --git a/profiler/src/main.cpp b/profiler/src/main.cpp
index 0b27109a..2c07a995 100644
--- a/profiler/src/main.cpp
+++ b/profiler/src/main.cpp
@@ -626,6 +626,9 @@ int main( int argc, char** argv )
             case tracy::LoadProgress::ContextSwitches:
                 ImGui::TextUnformatted( "Context switches..." );
                 break;
+            case tracy::LoadProgress::ContextSwitchesPerCpu:
+                ImGui::TextUnformatted( "CPU context switches..." );
+                break;
             default:
                 assert( false );
                 break;
diff --git a/server/TracyVersion.hpp b/server/TracyVersion.hpp
index fd4ff7b7..b8671501 100644
--- a/server/TracyVersion.hpp
+++ b/server/TracyVersion.hpp
@@ -7,7 +7,7 @@ namespace Version
 {
 enum { Major = 0 };
 enum { Minor = 5 };
-enum { Patch = 2 };
+enum { Patch = 3 };
 }
 }
 
diff --git a/server/TracyWorker.cpp b/server/TracyWorker.cpp
index b8870817..c8472c53 100644
--- a/server/TracyWorker.cpp
+++ b/server/TracyWorker.cpp
@@ -307,10 +307,14 @@ Worker::Worker( FileRead& f, EventType::Type eventMask )
     {
         s_loadProgress.total.store( 9, std::memory_order_relaxed );
     }
-    else
+    else if( fileVer <= FileVersion( 0, 5, 2 ) )
     {
         s_loadProgress.total.store( 10, std::memory_order_relaxed );
     }
+    else
+    {
+        s_loadProgress.total.store( 11, std::memory_order_relaxed );
+    }
 
     s_loadProgress.subTotal.store( 0, std::memory_order_relaxed );
     s_loadProgress.progress.store( LoadProgress::Initialization, std::memory_order_relaxed );
@@ -1429,6 +1433,44 @@ Worker::Worker( FileRead& f, EventType::Type eventMask )
         }
     }
 
+    if( fileVer >= FileVersion( 0, 5, 3 ) )
+    {
+        s_loadProgress.subTotal.store( 0, std::memory_order_relaxed );
+        s_loadProgress.progress.store( LoadProgress::ContextSwitchesPerCpu, std::memory_order_relaxed );
+        f.Read( sz );
+        s_loadProgress.subTotal.store( sz, std::memory_order_relaxed );
+        if( eventMask & EventType::ContextSwitches )
+        {
+            uint64_t cnt = 0;
+            for( int i=0; i<256; i++ )
+            {
+                int64_t refTime = 0;
+                f.Read( sz );
+                m_data.cpuData[i].cs.reserve_exact( sz, m_slab );
+                auto ptr = m_data.cpuData[i].cs.data();
+                for( uint64_t j=0; j<sz; j++ )
+                {
+                    ptr->SetStart( ReadTimeOffset( f, refTime ) );
+                    ptr->SetEnd( ReadTimeOffset( f, refTime ) );
+                    uint64_t thread;
+                    f.Read( thread );
+                    ptr->SetThread( thread );
+                    ptr++;
+                }
+                cnt += sz;
+                s_loadProgress.subProgress.store( cnt, std::memory_order_relaxed );
+            }
+        }
+        else
+        {
+            for( int i=0; i<256; i++ )
+            {
+                f.Read( sz );
+                f.Skip( sizeof( uint64_t ) * 3 * sz );
+            }
+        }
+    }
+
     s_loadProgress.total.store( 0, std::memory_order_relaxed );
     m_loadTime = std::chrono::duration_cast<std::chrono::nanoseconds>( std::chrono::high_resolution_clock::now() - loadStart ).count();
 
@@ -4826,6 +4868,24 @@ void Worker::Write( FileWrite& f )
             f.Write( &state, sizeof( state ) );
         }
     }
+
+    sz = GetContextSwitchPerCpuCount();
+    f.Write( &sz, sizeof( sz ) );
+    for( int i=0; i<256; i++ )
+    {
+        sz = m_data.cpuData[i].cs.size();
+        f.Write( &sz, sizeof( sz ) );
+        int64_t refTime = 0;
+        for( auto& cx : m_data.cpuData[i].cs )
+        {
+            WriteTimeOffset( f, refTime, cx.Start() );
+            WriteTimeOffset( f, refTime, cx.End() );
+            uint64_t thread = cx.Thread();
+            // Don't care about external thread identifiers
+            if( m_data.threadMap.find( thread ) == m_data.threadMap.end() ) thread = 0;
+            f.Write( &thread, sizeof( thread ) );
+        }
+    }
 }
 
 void Worker::WriteTimeline( FileWrite& f, const Vector<ZoneEvent*>& vec, int64_t& refTime )
diff --git a/server/TracyWorker.hpp b/server/TracyWorker.hpp
index fa8b1cc7..46f0429b 100644
--- a/server/TracyWorker.hpp
+++ b/server/TracyWorker.hpp
@@ -67,7 +67,8 @@ struct LoadProgress
         Memory,
         CallStacks,
         FrameImages,
-        ContextSwitches
+        ContextSwitches,
+        ContextSwitchesPerCpu
     };
 
     LoadProgress() : total( 0 ), progress( 0 ), subTotal( 0 ), subProgress( 0 ) {}