diff --git a/tinygrad/device.py b/tinygrad/device.py
index 49f4d020b1..37bbff4600 100644
--- a/tinygrad/device.py
+++ b/tinygrad/device.py
@@ -50,13 +50,13 @@ class JITRunner:
 def update_stats(name:str, op_estimate:sint, mem_estimate:sint, var_vals: Optional[Dict[Variable, int]], et: Optional[float], buf_count, jit=False, num_kernels=1, lra: Optional[Dict]=None):
   if var_vals is None: var_vals = {}
   op_estimate, mem_estimate = sym_infer(op_estimate, var_vals), sym_infer(mem_estimate, var_vals)
-  if DEBUG >= 2:
-    print(f"{colored(f'*** {GlobalCounters.kernel_count:4d}', ('magenta' if num_kernels == 1 else 'CYAN') if jit else None)} {name+' '*(37-ansilen(name))} arg {buf_count:3d} sz {str(lra.get('global_size', '') if lra else ''):18s} {str(lra.get('local_size', '') if lra else ''):12s} OPs {int(op_estimate/1e6):6d}M/{GlobalCounters.global_ops/1e9:7.2f}G  mem {GlobalCounters.mem_used/1e9:5.2f} GB " +
-          (str() if et is None else f"tm {et*1e6:9.2f}us/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({op_estimate/((et or 1e-20)*1e9):8.2f} GFLOPS, {mem_estimate/((et or 1e-20)*1e9):7.2f} GB/s)"))
   GlobalCounters.kernel_count += num_kernels
   GlobalCounters.global_ops += op_estimate
   GlobalCounters.global_mem += mem_estimate
   if et is not None: GlobalCounters.time_sum_s += et
+  if DEBUG >= 2:
+    print(f"{colored(f'*** {GlobalCounters.kernel_count:4d}', ('magenta' if num_kernels == 1 else 'CYAN') if jit else None)} {name+' '*(37-ansilen(name))} arg {buf_count:3d} sz {str(lra.get('global_size', '') if lra else ''):18s} {str(lra.get('local_size', '') if lra else ''):12s} OPs {int(op_estimate/1e6):6d}M/{GlobalCounters.global_ops/1e9:7.2f}G  mem {GlobalCounters.mem_used/1e9:5.2f} GB " +
+          (str() if et is None else f"tm {et*1e6:9.2f}us/{GlobalCounters.time_sum_s*1e3:9.2f}ms ({op_estimate/((et or 1e-20)*1e9):8.2f} GFLOPS, {mem_estimate/((et or 1e-20)*1e9):7.2f} GB/s)"))
 
 # **************** Buffer / Allocator ****************