Execution performance of OpenRAM

[FJShen]

Hi,

I was using OpenRAM for the first time, trying to generate an SRAM of 256kb size with FreePDK45 technology. It's already been running for 40 hours and still has not finished yet. Does it always take this long? Or am I doing something wrong?

This long run time really surprises me because there is a tutorial https://cornell-ece5745.github.io/ece5745-tut8-sram/ claiming it only takes a few minutes to generate a 4kb memory.

[mguthaus]

It shouldn’t but there could be instances where the power routing is difficult or characterization takes a long time. 256kb is getting big too for a single array but it should be do-able. What settings and version are you using?

the hard multi banking doesn’t work right now so for real big memories you need to wrap them with verilog banks. We have a “soft banking” feature being tested now for this.

[FJShen]

I'm using v1.1.19 (commit f66aac3). This is what my config looks like, looks like it is compiling for only one bank.

Technology: freepdk45
Total size: 262144 bits
Word size: 2048
Words: 128
Banks: 1
Write size: None
RW ports: 0
R-only ports: 1
W-only ports: 1

[mguthaus]

That’s an unusually large word size so I’ll have to check that out. It could be a runtime bug in our channel router.

[FJShen]

Ok thanks. FYI here's what the tool output to the terminal, including 2 warning messages.

Technology: freepdk45 Total size: 262144 bits Word size: 2048 Words: 128 Banks: 1 Write size: None RW ports: 0 R-only ports: 1 W-only ports: 1 DRC/LVS/PEX is only run on the top-level design to save run-time (inline_lvsdrc=True to do inline checking). DRC/LVS/PEX is disabled (check_lvsdrc=True to enable). Characterization is disabled (using analytical delay models) (analytical_delay=False to simulate). Only generating nominal corner timing. Words per row: 1 Output files are: ... Submodules: 643.4 seconds Placement: 61.2 seconds Retrieving pins: 0.3 seconds Analyzing pins: 0.2 seconds Finding blockages: 14374.0 seconds Converting blockages: 77.8 seconds WARNING: file pin_group.py: line 656: Expanding conversion (addr1[5] layer=m3 ll=v[2583.52,-3.8975] ur=v[2583.655,-3.7625]) WARNING: file pin_group.py: line 656: Expanding conversion (addr1[6] layer=m3 ll=v[2583.52,-6.1075] ur=v[2583.655,-5.9725]) Converting pins: 2417.9 seconds Separating adjacent pins: 60.0 seconds * Enclosing pins: 3.0 seconds Finding pins and blockages: 20003.0 seconds [stuck here so far]

[xobs]

Adding on to this, openram seems to take a very long time. sky130_sram_1kbyte_1r1w_8x1024_8.py has been running for an hour for me now. A process called magicdnull has 45 minutes of CPU time. Routing only took 815 seconds, but it's stopped there.

[mguthaus]

This is extraction in magic. It is slow...