diff --git a/crates/engine-gpu/src/lib.rs b/crates/engine-gpu/src/lib.rs
index 9a784aa..93bd20c 100644
--- a/crates/engine-gpu/src/lib.rs
+++ b/crates/engine-gpu/src/lib.rs
@@ -281,6 +281,14 @@ impl GpuEngine {
     pub fn device_count(&self) -> usize {
         self.contexts.len()
     }
+
+    /// Explicitly clear thread-local GPU resources.
+    /// Call this before thread exit to avoid TLS destruction order issues with wgpu.
+    pub fn clear_worker_resources() {
+        WORKER_RESOURCES.with(|resources| {
+            *resources.borrow_mut() = None;
+        });
+    }
 }
 
 impl MinerEngine for GpuEngine {
diff --git a/crates/miner-cli/Cargo.toml b/crates/miner-cli/Cargo.toml
index 2e1358c..c2bacd8 100644
--- a/crates/miner-cli/Cargo.toml
+++ b/crates/miner-cli/Cargo.toml
@@ -6,7 +6,7 @@ publish = false
 description = "CLI binary to run the Quantus External Miner service"
 
 [features]
-default = []
+default = ["gpu"]
 gpu = ["miner-service/gpu"]
 
 [dependencies]
diff --git a/crates/miner-cli/src/main.rs b/crates/miner-cli/src/main.rs
index e5eb261..e4c942b 100644
--- a/crates/miner-cli/src/main.rs
+++ b/crates/miner-cli/src/main.rs
@@ -402,8 +402,29 @@ async fn run_benchmark_command(
 
     // Chunk sizes
     let cpu_chunk = 10_000u64;
+    // Increase GPU chunk size to amortize submission overhead (10M)
+    // 1M might be too small for fast GPUs, 100M might be too large for short benchmarks.
     let gpu_chunk = 1_000_000u64;
 
+    if effective_gpu_workers > 0 {
+        if let Some(gpu_engine_arc) = &gpu_engine {
+             // We can't easily get device count from MinerEngine trait without downcasting or extending trait.
+             // But we know it's GpuEngine here if feature=gpu is enabled.
+             #[cfg(feature = "gpu")]
+             {
+                 if let Some(gpu_ptr) = gpu_engine_arc.as_any().downcast_ref::<engine_gpu::GpuEngine>() {
+                     let device_count = gpu_ptr.device_count();
+                     if effective_gpu_workers > device_count {
+                         println!("⚠️  WARNING: You have requested {} GPU workers for {} device(s).", effective_gpu_workers, device_count);
+                         println!("   Multiple workers per device creates contention and typically reduces hashrate.");
+                         println!("   Recommended: 1 worker per GPU device.");
+                         println!();
+                     }
+                 }
+             }
+        }
+    }
+
     for worker_id in 0..total_workers {
         // Determine type and engine
         let (engine, nonces_per_worker) = if worker_id < effective_cpu_workers {
@@ -487,6 +508,10 @@ async fn run_benchmark_command(
                 }
             }
 
+            // Cleanup TLS resources explicitly to avoid panic on exit
+            #[cfg(feature = "gpu")]
+            engine_gpu::GpuEngine::clear_worker_resources();
+
             worker_hashes
         });
 
diff --git a/crates/miner-service/Cargo.toml b/crates/miner-service/Cargo.toml
index ac8f33d..89b19b3 100644
--- a/crates/miner-service/Cargo.toml
+++ b/crates/miner-service/Cargo.toml
@@ -6,7 +6,7 @@ publish = false
 description = "Service layer: HTTP API, job orchestration, and engine abstraction for the Quantus External Miner"
 
 [features]
-default = ["cpu", "metrics"]
+default = ["cpu", "gpu", "metrics"]
 # Enable CPU engine by default.
 cpu = ["engine-cpu"]
 # Enable GPU engine.
diff --git a/crates/miner-service/src/lib.rs b/crates/miner-service/src/lib.rs
index 7bd18b4..49ace5d 100644
--- a/crates/miner-service/src/lib.rs
+++ b/crates/miner-service/src/lib.rs
@@ -933,6 +933,12 @@ fn mine_range_with_engine_typed(
         }
     }
 
+    // Explicitly clear thread-local GPU resources to avoid TLS order panic
+    #[cfg(feature = "gpu")]
+    if engine_type == "GPU" {
+        engine_gpu::GpuEngine::clear_worker_resources();
+    }
+
     log::debug!(target: "miner", "Job {job_id} thread {thread_id} completed.");
 }