1 lembach

A Diagnostic Method for Detectingand Assessing the Impact of Physical Design Optimizations on Routing : A Diagnostic Method for Detecting and Assessing the Impact of Physical Design Optimizations on Routing Robert Lembach Rafael A. Arce-Nazario Donald Eisenmenger Cory Wood IBM Engineering and Technology Services


Agenda : Agenda Appreciation Motivation and Goals Process Flow Examples Summary


Motivation – Improve Physical Design Quality : Motivation – Improve Physical Design Quality Serendipitous observations by physical designers using a variety of physical optimizations systems Poorly placed objects Sub-optimal buffer topologies or placements White space distribution issues Complexity: algorithms, versions, parameters, interactions Routing is being negatively impacted


What a Long Strange Trip It’s Been : What a Long Strange Trip It’s Been


Goals : Goals Enable an independent audit of physical designs from a variety of physical design systems Be exhaustive in scope Initial focus on rapidly increasing buffer quantities Easy to understand algorithm and metrics Enable data mining


Process Flow : Process Flow Interrogate net list to extract disjoint groups Execute algorithm on each group Data mining


Group Creation : Group Creation Groups can be serial and/or parallel buffering trees or other logic boxes. Groups are disjoint File is input for algorithm graphic tools data mining


Group Statistics for 8 Chip Designs : Group Statistics for 8 Chip Designs Chart includes only the transparent buffering cells Non-buffer groups also suitable for analysis


Group Colorization : Group Colorization From BlueGene/L chip (ISSCC 2005) Example of group use In a routing hotspot, find and move arbitrarily placed buffering to free up routing channels


OOB (Out of Bounds) Algorithm : OOB (Out of Bounds) Algorithm Compares original network to reduced network with buffering made transparent Calculated for each group Quality metrics Bloat length, ratio, density Laps around the chip


Data Mining: Meandering Buffer Chain : Data Mining: Meandering Buffer Chain Data mining technique Review 2-pin networks (buffering is transparent) OOB identifies this layout as grossly out of bounds with high bloat length and bloat ratio This area was hard to route


Data Mining: Tuning Fork Topology : Data Mining: Tuning Fork Topology Physical synthesis adds buffer near source to drive one of two far sinks. Far sinks are near each other. OOB predicts ~2x bloat, a doubling of routing demand Routing may be degraded if transform is repeated many times in local area


Data Mining: Tuning Fork Topology : Data Mining: Tuning Fork Topology Meandering nets reflect locally difficult routing OOB using actual routes shows andgt;2x bloat length One of several similar transforms in this area Timing surprises OOB can use estimated or actual routes


Data Mining: Non-buffer Groups : Data Mining: Non-buffer Groups OOB can be extended beyond buffered networks Example: 4-way OR with fanout of 1 on each net OOB predicts ~3X bloat length for this configuration For routing, better to fracture high function library elements, especially if they are locally clustered


Data Mining: One Bit in a Bus : Data Mining: One Bit in a Bus OOB detects high bloat length and ratio in simple buffer chain which one bit of a a larger bus Physical synthesis attempts to use holes punched in large object


Data Mining: All Bits : Data Mining: All Bits OOB detects issues wide variation in solution quality Physical synthesis attempts to randomly distribute the buffering Placement of buffering impacts routing, even if bloat is minimal


Data Mining: Placement Anomalies : Data Mining: Placement Anomalies


Data Mining: One-box OOB Groups : Data Mining: One-box OOB Groups Full chip view of bloat Objects can be displaced during legalization or overlap removal Addition of buffering is usually very non-uniform Useful in floor plan closure


Data Mining: Creative Buffering Schemes : Data Mining: Creative Buffering Schemes White object drives blue buffer and yellow objects Blue buffer drives red objects Blue buffer added to reduce load on white object Nearly doubles local wiring demand due to two nearly identical nets OOB: ~1.8x bloat ratio


Data Mining: Artificially Induced Problems : Data Mining: Artificially Induced Problems


Up to 10% of Chip Wire May Be Unnecessary : Up to 10% of Chip Wire May Be Unnecessary Buffer bloat (4%) Poor topology or poor placement of buffering Collateral Damage (4%) Proximate nets meandering due to added routing stress Proximate objects perturbed by buffer insertion Non-buffer bloat (2%) Library selection and influence on routing


Summary : Summary PD observations drove review of current practices Current tools do significant routing damage, with up to 10% of total chip wire unnecessary OOB flow is one way to track solution quality Data mining used to identify problems and trends