diff --git a/cuburn/render.py b/cuburn/render.py
index a7c754a..39f10fe 100644
--- a/cuburn/render.py
+++ b/cuburn/render.py
@@ -324,7 +324,7 @@ class RenderManager(ClsMod):
         # Split the launch into multiple rounds, possibly (slightly) reducing
         # work overlap but avoiding stalls when working on a device with an
         # active X session. TODO: characterize performance impact, autodetect
-        BLOCK_SIZE = 4
+        BLOCK_SIZE = 16
         for i in range(BLOCK_SIZE-1, nrounds, BLOCK_SIZE):
             launch_iter(BLOCK_SIZE)
         launch_iter(nrounds%BLOCK_SIZE)
diff --git a/dist/server.py b/dist/server.py
index a3a9eee..6159ba9 100644
--- a/dist/server.py
+++ b/dist/server.py
@@ -55,7 +55,7 @@ def setup_worker_listener(addrs, tq, rq):
             print ' >', ' '.join(addr)
             if task.hash not in compcache:
                 try:
-                    rsp = Renderer.compile(task.anim, arch='sm_21')
+                    rsp = Renderer.compile(task.anim, arch='sm_35')
                 except:
                     # Store exceptions, so that we don't try to endlessly
                     # recompile bad genomes