diff --git a/sumpy/p2p.py b/sumpy/p2p.py
index 7234180f62ff12d63b2e7c5afdaeaecf12c93f87..3071a3b88c9b1fc45c7d4ce3470535f084aba0a9 100644
--- a/sumpy/p2p.py
+++ b/sumpy/p2p.py
@@ -306,7 +306,12 @@ class P2PFromCSR(P2PBase):
     def get_optimized_kernel(self):
         # FIXME
         knl = self.get_kernel()
-        #knl = lp.split_iname(knl, "itgt_box", 16, outer_tag="g.0")
+        import pyopencl as cl
+        dev = self.context.devices[0]
+        if dev.type & cl.device_type.CPU:
+            knl = lp.split_iname(knl, "itgt_box", 4, outer_tag="g.0")
+        else:
+            knl = lp.split_iname(knl, "itgt_box", 4, outer_tag="g.0")
         return knl
 
     def __call__(self, queue, **kwargs):
diff --git a/sumpy/version.py b/sumpy/version.py
index 360986ffb5607fb5e4b69e29acfac449a232f76e..be874f65b3e22b32ad34f07f49670b51f8e32f5d 100644
--- a/sumpy/version.py
+++ b/sumpy/version.py
@@ -25,4 +25,4 @@ VERSION = (2016, 1)
 VERSION_STATUS = "beta1"
 VERSION_TEXT = ".".join(str(x) for x in VERSION) + VERSION_STATUS
 
-KERNEL_VERSION = 12
+KERNEL_VERSION = 13