diff --git a/src/nouveau/compiler/nak/assign_regs.rs b/src/nouveau/compiler/nak/assign_regs.rs
index 82de4014b12..644aeb587af 100644
--- a/src/nouveau/compiler/nak/assign_regs.rs
+++ b/src/nouveau/compiler/nak/assign_regs.rs
@@ -1486,6 +1486,28 @@ impl Shader<'_> {
             );
         }
 
+        // We try to check if allocating a few fewer registers would allow us
+        // to run 12 instead of 8 warps concurrently.
+        // The assumption is that running +50% threads will give us more
+        // performance than we lose with a bit of spilling.
+        let actual_gprs = (total_gprs + hw_reserved_gprs).min(max_gprs);
+        let warps_per_sm = max_warps_per_sm(self.sm, actual_gprs);
+        if warps_per_sm == 8 {
+            let new_max = if max_warps_per_sm(self.sm, actual_gprs - 8) > 8 {
+                total_gprs - 8
+            } else if max_warps_per_sm(self.sm, actual_gprs - 16) > 8 {
+                // This gives us +15% performance in pixmark_piano
+                total_gprs - 16
+            } else {
+                0
+            };
+
+            if new_max != 0 {
+                max_gprs = (new_max.next_multiple_of(8) - hw_reserved_gprs)
+                    .min(max_gprs);
+            }
+        }
+
         if total_gprs > max_gprs {
             // If we're spilling GPRs, we need to reserve 2 GPRs for OpParCopy
             // lowering because it needs to be able lower Mem copies which