Updated on 2024-08-18

Author: lxaw

index.html

@@ -39,7 +39,7 @@ <h3>
         When?
     </h3>
     <p>
-        Last time this was edited was 2024-08-13 (YYYY/MM/DD).
+        Last time this was edited was 2024-08-18 (YYYY/MM/DD).
     </p>
     <small><a href="misc.html">misc</a></small>
 </body>

papers/list.json

@@ -5,7 +5,7 @@
     "year": "2018",
     "topic": "quantization, floating-point, precision",
     "venue": "NeurIps",
-    "description": "The authors show that it is possible to train DNNs with 8-bit fp values while maintaining decent accuracy. To do this, they make a new FP8 format, develope a technique \"chunk-based computations\" that allow matrix and convolution ops to be computed using 8-bit multiplications and 16 bit additions, and use fp stochastic rounding in weight updates.",
+    "description": "The authors show that it is possible to train DNNs with 8-bit fp values while maintaining decent accuracy. To do this, they make a new FP8 format, develope a technique \"chunk-based computations\" that allow matrix and convolution ops to be computed using 8-bit multiplications and 16 bit additions, and use fp stochastic rounding in weight updates. One interesting point they make is that swamping (the issue of truncation in large-to-small number addition) is a serious problem in DNN bit-precision reduction.",
     "link": "https://arxiv.org/pdf/1812.08011"
   },
   {