diff --git a/tinygrad/runtime/ops_amd.py b/tinygrad/runtime/ops_amd.py
index e5f968133d..3a30cb6791 100644
--- a/tinygrad/runtime/ops_amd.py
+++ b/tinygrad/runtime/ops_amd.py
@@ -25,6 +25,7 @@ SQTT = ContextVar("SQTT", abs(VIZ.value)>=2)
 SQTT_ITRACE_SE_MASK, SQTT_LIMIT_SE, SQTT_SIMD_SEL, SQTT_TOKEN_EXCLUDE = \
   ContextVar("SQTT_ITRACE_SE_MASK", 0b11), ContextVar("SQTT_LIMIT_SE", 0), ContextVar("SQTT_SIMD_SEL", 0), ContextVar("SQTT_TOKEN_EXCLUDE", 0)
 PMC = ContextVar("PMC", abs(VIZ.value)>=2)
+AMD_USB_POLL_US = getenv("AMD_USB_POLL_US", 500) # microseconds to sleep between USB signal polls. 0 disables
 EVENT_INDEX_PARTIAL_FLUSH = 4 # based on a comment in nvd.h
 WAIT_REG_MEM_FUNCTION_EQ  = 3 # ==
 WAIT_REG_MEM_FUNCTION_NEQ = 4 # !=
@@ -45,8 +46,9 @@ class AMDSignal(HCQSignal):
   def __init__(self, *args, **kwargs): super().__init__(*args, **{**kwargs, 'timestamp_divider': 100})
 
   def _sleep(self, time_spent_since_last_sleep_ms:int):
-    # Reasonable to sleep for long workloads (which take more than 200ms) and only timeline signals.
-    if time_spent_since_last_sleep_ms > 200 and self.owner is not None: self.owner.iface.sleep(200)
+    # USB signals live in VRAM across the link, so yield between polls. Native AMD only blocks after 200 ms.
+    if self.owner is not None and self.owner.is_usb() and AMD_USB_POLL_US: time.sleep(AMD_USB_POLL_US / 1e6)
+    elif time_spent_since_last_sleep_ms > 200 and self.owner is not None: self.owner.iface.sleep(200)
 
 class AMDComputeQueue(HWQueue):
   def __init__(self, dev:AMDDevice):