From 6e20e640f1e9b0e985e3e0d224aa5504ccf0699c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Matthias=20K=C3=B6fferlein?= Date: Wed, 6 Feb 2019 01:46:36 +0100 Subject: [PATCH] Updated 2019 02 05 (markdown) --- 2019-02-05.md | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/2019-02-05.md b/2019-02-05.md index 3f3179d..115de56 100644 --- a/2019-02-05.md +++ b/2019-02-05.md @@ -207,8 +207,10 @@ Deep cont&poly ... 0.01s (sys), 1.25s (user), 1.265s (wall) Deep cont-poly ... 0s (sys), 1.31s (user), 1.306s (wall) ``` -So overall a great improvement in the hierarchical frontend layers (factor 5 to 20 in layers 1-10), a roughly factor 2 in layer 11 (metal 1) which has both hierarchical parts in the local wiring of the standard cells, but also flat parts from the power routing and pins. There is a 2-3x disadvantage of the hierarchical implementation in the flat layers (14-26) because the hierarchical analysis demands some overhead and there is no gain for the merge computation itself. +So overall a great improvement in the hierarchical frontend layers (factor 5 to 20 in layers 1-10). Still about factor 2 in layer 11 (metal 1) which has both hierarchical parts in the local wiring of the standard cells, but also flat parts from the power routing and pins. There is a 2-3x disadvantage of the hierarchical implementation in the flat layers (14-26) because the hierarchical analysis demands some overhead and there is no gain for the merge computation itself. But the booleans pay this effort back with a 8x performance boost. That's because they are mainly frontend-driven and benefit from the hierarchical nature of the standard cells. - +Total runtime of the script is in my case + * 46s (hierarchical) + * 210s (flat)