More orojenesis interperetation

tanner-andrulis · tanner-andrulis · commit 9edac0a5fc08 · 2026-02-16T12:06:52.000-05:00
diff --git a/notebooks/tutorials/memory_size_vs_access_tradeoff.ipynb b/notebooks/tutorials/memory_size_vs_access_tradeoff.ipynb
@@ -182,7 +182,11 @@
     "\n",
     "The curve slowly flattens out, rather than quickly, because each Einsum has a different\n",
     "shape, and some Einsums enter the no-refetch, fusion-helpful region at smaller on-chip\n",
-    "memory sizes."
+    "memory sizes.\n",
+    "\n",
+    "Finally, there is a steep drop-off for the fused case at larger on-chip memory sizes.\n",
+    "This occurs when it becomes feasible to fuse all attention Einsums at once, giving a\n",
+    "substantial reduction in off-chip accesses."
    ]
   },
   {

Original file line number	Diff line number	Diff line change
`@@ -182,7 +182,11 @@`
`182`	`182`	`"\n",`
`183`	`183`	`"The curve slowly flattens out, rather than quickly, because each Einsum has a different\n",`
`184`	`184`	`"shape, and some Einsums enter the no-refetch, fusion-helpful region at smaller on-chip\n",`
`185`		`- "memory sizes."`
	`185`	`+ "memory sizes.\n",`
	`186`	`+ "\n",`
	`187`	`+ "Finally, there is a steep drop-off for the fused case at larger on-chip memory sizes.\n",`
	`188`	`+ "This occurs when it becomes feasible to fuse all attention Einsums at once, giving a\n",`
	`189`	`+ "substantial reduction in off-chip accesses."`
`186`	`190`	`]`
`187`	`191`	`},`
`188`	`192`	`{`