diff --git a/src/common/thread_worker.h b/src/common/thread_worker.h
index ec3ca0af49..a88713665d 100644
--- a/src/common/thread_worker.h
+++ b/src/common/thread_worker.h
@@ -117,6 +117,13 @@ class StatefulThreadWorker {
         });
     }
 
+    // Drops tasks no worker has started. Running tasks still finish.
+    void ClearRequests() {
+        std::scoped_lock lock{queue_mutex};
+        work_scheduled -= requests.size();
+        requests = std::queue<Task>{};
+    }
+
 private:
     std::queue<Task> requests;
     std::mutex queue_mutex;
diff --git a/src/shader_recompiler/backend/spirv/emit_spirv_mesh.cpp b/src/shader_recompiler/backend/spirv/emit_spirv_mesh.cpp
index 6c599528cc..e0f3f437f6 100644
--- a/src/shader_recompiler/backend/spirv/emit_spirv_mesh.cpp
+++ b/src/shader_recompiler/backend/spirv/emit_spirv_mesh.cpp
@@ -207,6 +207,13 @@ void EmitIsbeWriteMapU32(EmitContext& ctx, const IR::Value& address, Id value) {
 }
 
 void EmitMeshPrologue(EmitContext& ctx) {
+    if (ctx.stage == Stage::Task) {
+        // A task that culls all its threads never writes the mesh workgroup count, and the payload
+        // starts undefined. Intel's garbage there launches enough workgroups to lose the device.
+        ctx.OpStore(PayloadWord(ctx, ctx.Const(1U)), ctx.u32_zero_value);
+        Barrier(ctx);
+        return;
+    }
     if (ctx.stage != Stage::Mesh) {
         return;
     }
diff --git a/src/shader_recompiler/ir_opt/conditional_barrier_pass.cpp b/src/shader_recompiler/ir_opt/conditional_barrier_pass.cpp
index c3ed27f4f2..8a792b2ff2 100644
--- a/src/shader_recompiler/ir_opt/conditional_barrier_pass.cpp
+++ b/src/shader_recompiler/ir_opt/conditional_barrier_pass.cpp
@@ -31,6 +31,7 @@ void ConditionalBarrierPass(IR::Program& program) {
                     inst.GetOpcode() == IR::Opcode::Barrier) {
                     LOG_WARNING(Shader, "Barrier within conditional control flow");
                     inst.ReplaceOpcode(IR::Opcode::Identity);
+                    program.info.removed_conditional_barrier = true;
                 }
             }
             break;
diff --git a/src/shader_recompiler/shader_info.h b/src/shader_recompiler/shader_info.h
index 4e5fb5c6e4..d2e2c807e0 100644
--- a/src/shader_recompiler/shader_info.h
+++ b/src/shader_recompiler/shader_info.h
@@ -320,6 +320,7 @@ struct Info {
     bool uses_sparse_residency{};
     bool uses_demote_to_helper_invocation{};
     bool uses_subgroup_vote{};
+    bool removed_conditional_barrier{};
     bool uses_subgroup_mask{};
     bool uses_fswzadd{};
     bool uses_derivatives{};
diff --git a/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp b/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp
index ae008b48a7..5fa13eb0fa 100644
--- a/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp
+++ b/src/video_core/renderer_vulkan/vk_compute_pipeline.cpp
@@ -88,7 +88,11 @@ ComputePipeline::ComputePipeline(const Device& device_, Scheduler& scheduler, vk
         };
         // Only warp-level ops observe the subgroup size. Forcing 32 on other shaders makes Intel
         // compile SIMD32, which can spill a register-heavy kernel for minutes (Mario Kart World).
-        const bool uses_warp_ops = info.uses_subgroup_vote || info.uses_subgroup_invocation_id ||
+        // A shader that lost a conditional barrier relies on a whole warp running in step: at a
+        // smaller subgroup size it reads stale shared memory, which shows as banding over the image
+        // in Mario Kart World on Intel's Windows driver.
+        const bool uses_warp_ops = info.removed_conditional_barrier ||
+                                   info.uses_subgroup_vote || info.uses_subgroup_invocation_id ||
                                    info.uses_subgroup_shuffles || info.uses_subgroup_mask ||
                                    info.uses_quad_shuffles || info.uses_fswzadd;
         VkPipelineCreateFlags flags{};
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
index 18aa74b29e..e926a0a1f8 100644
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
@@ -661,6 +661,11 @@ bool PreloadAllPipelines() {
 PipelineCache::~PipelineCache() {
     SaveUsedPipelines();
     if (use_vulkan_pipeline_cache && !vulkan_pipeline_cache_filename.empty()) {
+        if (!device.CanReadPipelineCacheDuringBuilds()) {
+            // Drop queued builds, as destroying the workers would; running ones must end first.
+            workers.ClearRequests();
+            workers.WaitForRequests();
+        }
         SerializeVulkanPipelineCache(vulkan_pipeline_cache_filename, vulkan_pipeline_cache,
                                      CACHE_VERSION);
     }
@@ -912,7 +917,8 @@ void PipelineCache::LoadDiskResources(u64 title_id, std::stop_token stop_loading
 }
 
 void PipelineCache::QueueVulkanPipelineCacheFlush() {
-    if (!use_vulkan_pipeline_cache || vulkan_pipeline_cache_filename.empty()) {
+    if (!use_vulkan_pipeline_cache || vulkan_pipeline_cache_filename.empty() ||
+        !device.CanReadPipelineCacheDuringBuilds()) {
         return;
     }
     if (++pipelines_since_flush < VULKAN_CACHE_FLUSH_PIPELINES) {
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index bb8c7f57a3..f51182f044 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -246,6 +246,8 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
 
 RasterizerVulkan::~RasterizerVulkan() {
     VideoCommon::GpuSplit::Get().SetIdleHook(nullptr, nullptr);
+    // The channels are gone by now, and submitting invalidates the state tracker's flags.
+    state_tracker.ClearChannel();
     scheduler.WaitWorker();
     scheduler.Finish();
 }
@@ -2553,9 +2555,13 @@ void RasterizerVulkan::BindChannel(Tegra::Control::ChannelState& channel) {
     query_cache.BindToChannel(channel_id);
     state_tracker.ChangeChannel(channel);
     state_tracker.InvalidateState();
+    bound_channel_id = channel_id;
 }
 
 void RasterizerVulkan::ReleaseChannel(s32 channel_id) {
+    if (channel_id == bound_channel_id) {
+        state_tracker.ClearChannel();
+    }
     EraseChannel(channel_id);
     {
         std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.h b/src/video_core/renderer_vulkan/vk_rasterizer.h
index 1e550a27ca..65de3105d2 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.h
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.h
@@ -224,6 +224,7 @@ class RasterizerVulkan final : public VideoCore::RasterizerInterface,
     const Device& device;
     MemoryAllocator& memory_allocator;
     StateTracker& state_tracker;
+    s32 bound_channel_id = -1;
     Scheduler& scheduler;
 
     StagingBufferPool staging_pool;
diff --git a/src/video_core/renderer_vulkan/vk_state_tracker.h b/src/video_core/renderer_vulkan/vk_state_tracker.h
index 6b47ba4176..d47fa9e0e2 100644
--- a/src/video_core/renderer_vulkan/vk_state_tracker.h
+++ b/src/video_core/renderer_vulkan/vk_state_tracker.h
@@ -290,6 +290,11 @@ class StateTracker {
 
     void ChangeChannel(Tegra::Control::ChannelState& channel_state);
 
+    /// Stops using the bound channel's flags, for when it goes away.
+    void ClearChannel() {
+        flags = &default_flags;
+    }
+
     void InvalidateState();
 
 private:
diff --git a/src/video_core/vulkan_common/vulkan_device.h b/src/video_core/vulkan_common/vulkan_device.h
index f3f9209c99..c69edb6d2d 100644
--- a/src/video_core/vulkan_common/vulkan_device.h
+++ b/src/video_core/vulkan_common/vulkan_device.h
@@ -369,6 +369,12 @@ class Device {
                id != VK_DRIVER_ID_MESA_RADV;
     }
 
+    /// Whether a pipeline cache can be read while pipelines build into it, as Vulkan allows. Intel's
+    /// Windows driver crashes in vkGetPipelineCacheData when a build is running.
+    bool CanReadPipelineCacheDuringBuilds() const {
+        return GetDriverID() != VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS;
+    }
+
     /// Returns true for tile-based deferred renderers.
     bool IsTiler() const {
         switch (GetDriverID()) {