[MacroPlacement](../../) is an open, transparent effort to provide a public, baseline implementation of [Google Brain’s Circuit Training](https://github.com/google-research/circuit_training)(Morpheus) deep RL-based placement method. In this repo, we aim to achieve the following.
[MacroPlacement](../../) is an open, transparent effort to provide a public, baseline implementation of [Google Brain’s Circuit Training](https://github.com/google-research/circuit_training)(Morpheus) deep RL-based placement method. In this repo, we aim to achieve the following.
...
@@ -4214,7 +4214,9 @@ We have trained CT to generate a macro placement for the [MemPool Group design](
...
@@ -4214,7 +4214,9 @@ We have trained CT to generate a macro placement for the [MemPool Group design](
**November 25:**
**November 25:**
<aid="November25"></a>
<aid="November25"></a>
We document two variant Evaluation Flows (taking macro placements through Innovus place-and-route) that we use, in this [Evaluation Flow document](https://docs.google.com/document/d/1xDGFSYxIE0AKsGAI3ccLz1EX3bLHOvDtwl3983G5kYk/edit?usp=sharing). Posted results up to now have been obtained with Evaluation Flow 2. The [Evaluation Flow document](https://docs.google.com/document/d/1xDGFSYxIE0AKsGAI3ccLz1EX3bLHOvDtwl3983G5kYk/edit?usp=sharing) shows that results and conclusions are nearly identical between Evaluation Flow 1 and Evaluation Flow 2. However, going forward we will report our macro placement assessments using Evaluation Flow 1.
**CT Results for Commercial Enablement**
We have run CT to generate macro placement for Ariane133, BlackParrot and MemPool Group designs on GLOBALFOUNDRIES 12nm (GF12) enablement. The following tables present the normalized design metrics. Core area, standard cell area and macro area are normalized with respect to the core area. Total power is normalized w.r.t. the reported preCTS total power when CMP is used. Similarly, we normalize the wirelength and congestion based on the reported preCTS wirelength and congestion when CMP is used. The timing numbers are normalized w.r.t. the target clock period.
We have run CT to generate macro placement for Ariane133, BlackParrot and MemPool Group designs on GLOBALFOUNDRIES 12nm (GF12) enablement. The following tables present the normalized design metrics. Core area, standard cell area and macro area are normalized with respect to the core area. Total power is normalized w.r.t. the reported preCTS total power when CMP is used. Similarly, we normalize the wirelength and congestion based on the reported preCTS wirelength and congestion when CMP is used. The timing numbers are normalized w.r.t. the target clock period.
- The following table and screenshots provide details of Ariane133 GF12 implementation when CMP is used to generate the initial macro placement.
- The following table and screenshots provide details of Ariane133 GF12 implementation when CMP is used to generate the initial macro placement.
### Macro placement generated by an industry expert (The canvas is not gridded)
### Macro placement generated by an industry expert (The canvas is not gridded)
We thank Dr. Jinwook Jung for sharing the Human Macro Placement of Ariane133 design. The following figure shows placed and routed Ariane133-NG45 design where the macro palcement is generated by Dr. Jung. He has also shared the[place_srams.tcl](./def/place_srams.tcl) to reproduce the macro placement.
We thank Dr. Jinwook Jung of IBM Research for providing his Human Macro Placement of Ariane133 design as an alternative baseline. The following figure shows the placed and routed Ariane133-NG45 design, where the macro placement is generated by Dr. Jung. Dr. Jung has also provided his[place_srams.tcl](./def/place_srams.tcl) to reproduce the macro placement.
