diff --git a/tinygrad/runtime/graph/hcq.py b/tinygrad/runtime/graph/hcq.py
index ec62d829c0..868d1cc5d7 100644
--- a/tinygrad/runtime/graph/hcq.py
+++ b/tinygrad/runtime/graph/hcq.py
@@ -50,7 +50,7 @@ class HCQGraph(MultiGraphRunner):
 
     self.comp_queues: dict[HCQCompiled, HWQueue] = {dev: dev.hw_compute_queue_t() for dev in self.devices}
     self.copy_queues: dict[tuple[HCQCompiled, int], HWQueue] = {} # lazy allocation, keyed by (device, queue_idx)
-    self.num_copy_queues: int = getenv("HCQ_NUM_SDMA", 2 if ALL2ALL >= 1 else 1)
+    self.num_copy_queues: int = getenv("HCQ_NUM_SDMA", 7 if ALL2ALL >= 1 else 1)
     self.copy_queue_cnt: collections.defaultdict[HCQCompiled, itertools.count] = collections.defaultdict(itertools.count)
 
     self.signals: dict[Any, HCQSignal] = {**{dev: dev.new_signal(value=0) for dev in self.devices if not dev._is_cpu()},
diff --git a/tinygrad/runtime/support/am/ip.py b/tinygrad/runtime/support/am/ip.py
index 1f01a2389d..97769638a3 100644
--- a/tinygrad/runtime/support/am/ip.py
+++ b/tinygrad/runtime/support/am/ip.py
@@ -453,7 +453,7 @@ class AM_SDMA(AM_IP):
       self.adev.regGRBM_SOFT_RESET.write(0x0)
 
   def setup_ring(self, ring_addr:int, ring_size:int, rptr_addr:int, wptr_addr:int, idx:int) -> tuple[int, int]:
-    pipe, queue = idx % 4, idx // 4
+    pipe, queue = idx // 4, idx % 4
     reg, inst = ("regSDMA_GFX", pipe+queue*4) if self.adev.ip_ver[am.SDMA0_HWIP][:2] == (4,4) else (f"regSDMA{pipe}_QUEUE{queue}", 0)
     doorbell = am.AMDGPU_NAVI10_DOORBELL_sDMA_ENGINE0 + (pipe+queue*4) * 0xA
     self.sdma_reginst.append((reg, inst))