diff --git a/profiler/src/profiler/TracyView_ZoneInfo.cpp b/profiler/src/profiler/TracyView_ZoneInfo.cpp
index a466bf2c..0f5ff3c1 100644
--- a/profiler/src/profiler/TracyView_ZoneInfo.cpp
+++ b/profiler/src/profiler/TracyView_ZoneInfo.cpp
@@ -2025,12 +2025,6 @@ void View::ZoneTooltip( const GpuEvent& ev )
         }
     }
 
-    ImGui::Separator();
-    TextDisabledUnformatted( "GPU zone debug" );
-    TextFocused( "GPU begin (Tracy ns):", RealToString( ev.GpuStart() ) );
-    TextFocused( "GPU end (Tracy ns):", RealToString( end ) );
-    TextFocused( "Begin query id:", RealToString( ev.query_id ) );
-
     ImGui::EndTooltip();
 }
 
diff --git a/public/client/TracyProfiler.cpp b/public/client/TracyProfiler.cpp
index a44e255e..7d527517 100644
--- a/public/client/TracyProfiler.cpp
+++ b/public/client/TracyProfiler.cpp
@@ -84,8 +84,6 @@
 #include "TracySysTrace.hpp"
 #include "../tracy/TracyC.h"
 
-#include "../tracy/Tracy.hpp"
-
 #if defined TRACY_MANUAL_LIFETIME && !defined(TRACY_DELAYED_INIT)
 #  error "TRACY_MANUAL_LIFETIME requires enabled TRACY_DELAYED_INIT"
 #endif
diff --git a/server/TracyEvent.hpp b/server/TracyEvent.hpp
index d8dbad88..075479a5 100644
--- a/server/TracyEvent.hpp
+++ b/server/TracyEvent.hpp
@@ -380,19 +380,9 @@ struct GpuEvent
     tracy_force_inline int64_t CpuEnd() const { return int64_t( _cpuEnd_thread ) >> 16; }
     tracy_force_inline void SetCpuEnd( int64_t cpuEnd ) { assert( cpuEnd < (int64_t)( 1ull << 47 ) ); memcpy( ((char*)&_cpuEnd_thread)+2, &cpuEnd, 4 ); memcpy( ((char*)&_cpuEnd_thread)+6, ((char*)&cpuEnd)+4, 2 ); }
     tracy_force_inline int64_t GpuStart() const { return int64_t( _gpuStart_child1 ) >> 16; }
-    tracy_force_inline void SetGpuStart( int64_t gpuStart )
-    {
-        assert( gpuStart < (int64_t)( 1ull << 47 ) );
-        memcpy( ((char*)&_gpuStart_child1)+2, &gpuStart, 4 );
-        memcpy( ((char*)&_gpuStart_child1)+6, ((char*)&gpuStart)+4, 2 );
-    }
+    tracy_force_inline void SetGpuStart( int64_t gpuStart ) { /*assert( gpuStart < (int64_t)( 1ull << 47 ) );*/ memcpy( ((char*)&_gpuStart_child1)+2, &gpuStart, 4 ); memcpy( ((char*)&_gpuStart_child1)+6, ((char*)&gpuStart)+4, 2 ); }
     tracy_force_inline int64_t GpuEnd() const { return int64_t( _gpuEnd_child2 ) >> 16; }
-    tracy_force_inline void SetGpuEnd( int64_t gpuEnd )
-    {
-        assert( gpuEnd < (int64_t)( 1ull << 47 ) );
-        memcpy( ((char*)&_gpuEnd_child2)+2, &gpuEnd, 4 );
-        memcpy( ((char*)&_gpuEnd_child2)+6, ((char*)&gpuEnd)+4, 2 );
-    }
+    tracy_force_inline void SetGpuEnd( int64_t gpuEnd ) { assert( gpuEnd < (int64_t)( 1ull << 47 ) ); memcpy( ((char*)&_gpuEnd_child2)+2, &gpuEnd, 4 ); memcpy( ((char*)&_gpuEnd_child2)+6, ((char*)&gpuEnd)+4, 2 ); }
     tracy_force_inline int16_t SrcLoc() const { return int16_t( _cpuStart_srcloc & 0xFFFF ); }
     tracy_force_inline void SetSrcLoc( int16_t srcloc ) { memcpy( &_cpuStart_srcloc, &srcloc, 2 ); }
     tracy_force_inline uint16_t Thread() const { return uint16_t( _cpuEnd_thread & 0xFFFF ); }
diff --git a/server/TracyWorker.cpp b/server/TracyWorker.cpp
index 7821899c..21bd21ba 100644
--- a/server/TracyWorker.cpp
+++ b/server/TracyWorker.cpp
@@ -5924,7 +5924,7 @@ void Worker::ProcessGpuZoneEnd( const QueueGpuZoneEnd& ev, bool serial )
     assert( !td->second.stack.empty() );
     auto zone = td->second.stack.back_and_pop();
 
-    assert( !ctx->query[ev.queryId] ); 
+    assert( !ctx->query[ev.queryId] );
     ctx->query[ev.queryId] = zone;
 
     int64_t cpuTime;
@@ -5946,16 +5946,8 @@ void Worker::ProcessGpuTime( const QueueGpuTime& ev )
     auto ctx = m_gpuCtxMap[ev.context];
     assert( ctx );
 
-    if (ev.gpuTime < -1'000'000'000)
-        __debugbreak();
-
     int64_t tgpu = RefTime( m_refTimeGpu, ev.gpuTime );
-
-    // This overflow logic is for detecting "backward time jumps" due to asynchronous
-    // GPU scheduling (commands being rescheduled internally by the command processor,
-    // instead of being forced to execute in submit queue order)
-    constexpr int64_t gracePeriod = ( 1u << 31 ); // 2^31 ticks (about 2s, when 1 tick = 1ns)
-    if( tgpu < ctx->lastGpuTime - gracePeriod )
+    if( tgpu < ctx->lastGpuTime - ( 1u << 31 ) )
     {
         if( ctx->overflow == 0 )
         {
@@ -5963,6 +5955,7 @@ void Worker::ProcessGpuTime( const QueueGpuTime& ev )
         }
         ctx->overflowMul++;
     }
+    ctx->lastGpuTime = tgpu;
     if( ctx->overflow != 0 )
     {
         tgpu += ctx->overflow * ctx->overflowMul;