Semi-Custom Design Flow: Leveraging Place and Route Tools in Custom Circuit Design

Nadeem Eleyan Ken Lin Masud Kamal Baker Mohammad Paul Bassett [email protected] [email protected] [email protected] [email protected] [email protected]

Outline

Motivation for semi-custom Flow overview Usage model Better visibility Flow details Examples Conclusion

Introduction

IC designers have two options to implement a circuit block:

Synthesis / Auto place and route (ASIC) Custom circuit design / Custom Layout (Full Custom) Design complexity Timing requirements Area requirements Power requirements Project Schedule and Resources Block containing some non-static circuits (SRAM or dynamic) end up as Full Custom In reality only a portion of that block is non-static Most blocks also contain some standard CMOS circuits (data path and control logic) These portions can be built as ASIC or tiled standard cells.

Choice is based on the following:

Problem: designers tend to think of a hard boundary between the two flows:

Goal: allow designers to mix and match aspect from both ASIC and Full Custom approaches to improve productivity

Full-Custom vs. Semi-Custom vs. ASIC

Semi-Custom flow overview

The Semi-Custom block partitioning:

True Custom Macro Sub-block:
Smallest portion of design that has to be Full Custom Ex: SRAM array with non-static, small swing circuits Primary interface to sub-block is fully static Place and Route unit using standard Place and Route flow Uses custom pitch matching, tiling and off grid pre-routing

Soft Macro Wrapper:

Different portions of design handled in different ways:

SRAM and non-static periphery still Full Custom Non-critical logic handled as standard ASIC Timing critical data-paths tiled and pitch-matched to SRAM

Semi-Custom flow usage model

Problem: Full Custom design has long iteration time and upfront planning:
Circuit designer has to plan every single detail of the block Mask designer has to draw every polygon of the layout before we can have fully routed design First pass:
Create a simple floor plan Only pre-place True Custom Sub-Blocks Let the standard place and route tool finish off the design

Semi-custom flow uses an iterative approach:

First pass normally yields bad timing and routing results, but is used as a reference point Next Iteration:
Tile and/or pre-route top critical portions of the design Dont have to address all the critical paths at once

Since iteration time is short (few hours) we can have a fully routed first pass design very quickly Keep iterating until acceptable results are reached

At any point during this process we can stop and have a fully routed design Result: better trade off between how much to optimize vs. how quick to finish Extreme usage case:
Manually size and pre-place each cell in the design Pre-route each net. Both results and effort will be comparable to a Full Custom Block.

Better Visibility

Top level analysis flows (Timing / Power / Noise) have more visibility into Semi-Custom Blocks Top level:
Uses gate level tools (PrimeTime, BlastFusion, Talus, RedHawk) Block level analyzed with transistor level tools (HSpice, HSim , Nanotime, Totum) Black box Timing / Noise / Power / Physical Abstract

Full Custom Blocks:

Black boxing can cause miscommunication and inaccuracies Semi-Custom blocks allow top level visibility down to standard cells and custom sub-block Abstraction still needed for True Custom Macro Sub block That portion of design is much smaller and can be analyzed more easily

Semi-Custom flow details

Semi-Custom flow is an Auto Place & Route flow with additional hooks:
Force-Keep cells and nets Custom tiling Custom pre-routing
Cells to be tiled must have predictable names Nets to be pre-routed must have predictable names

Special netlist requirements:

Synthesis flows do not guarantee these conditions Methods to create the netlist:
RTL macros expand to predetermined gate level structures Write netlist manually by hand or script Schematic entry using standard cells

We chose the schematic entry method because our designers were more comfortable with it

Semi-Custom flow details (cont.)

Force-Keep Cells & Nets:

Default behavior of standard Place and Route flows is to optimize any gate or net in the design Problem: tiling scripts may fail if instances disappear or change

Same is true for nets we intend to pre-route if the flow inserts buffers in them Solution: mark cells to be tiled and nets to be pre-routed with Force Keep This ensures that they are still in the netlist when we reach the tiling and routing stages This however does not prevent the flow from upsizing / downsizing the gates as need

Semi-Custom flow details (cont.)

Custom Tiling:
Tile critical parts of design Use different algorithms depending on the context Ex: Tile cells in reference to a custom sub blocks pins Diagram shows typical pitch matching/tiling example:
Left side has Memory array with non-standard pitch Tile next stage of logic to minimizes vertical routing Place each cell in the same row as the custom pin it needs to be routed to Both Inv<0> and Inv<1> need to be in second row Use collision detections code to legalizes locations

Might end up with collisions:

10

Semi-Custom flow details (cont.)

Custom Pre-Routing:
Pre-route critical nets in the design Standard auto router is timing driven:
Tends to give certain bits of a regular structure higher priority than others This results in non-uniform routing and congestion

Pre-routing ensure uniformity and congestion relief Pre-routing also used to guide router through off-grid routing resource in sub-arrays:
net<0:3> needs to be routed through the sub-array Each bit has only one metal 4 off-grid open track available Pre-route metal 4 wires over the sub-array Run standard router to finish off the route

11

Semi-Custom Examples

Semi-Custom flow successfully use in Qualcomm high performance DSP core in 45nm technology The two areas the flow was used:
Memory blocks tiling Pure data-path tiling

Memory block tiling example:

4 K Bit, 6 Read, 4 Write multi-port register file. four 1K Full Custom sub arrays Tiled first stage muxing between sub-arrays Pre-routed mux outputs through sub-arrays

Tiled design
12

Semi-Custom Examples / Register file

Pre-routed design

Fully placed and routed design

13

Semi-Custom Examples / Register file (cont.)

The goal of the tiling and pre-routing in this case was to guide the Auto Place & Route flow out of the congested region between the sub arrays.

Congestion map w/o tiling

Congestion map with tiling

14

Semi-Custom Examples / Register file (cont.)

no tiling or pre-routing Total cell count Total cell area Total wire count Total wire length Cells tiled Wires pre-routed Short/Open/DRC 10120 0.042 mm2 140946 0.92459 m 0 (0%) 0 (0%) 339/42/135 10923 0.039 mm2 174025 1.01119 m 1536 (14%) 0 (0%) 0/0/0 with tiling only with tiling & prerouting

10910 0.038 mm2 139375 0.80323 m 1536 (14%) 3388 (3%) 0/0/0

Table shows the following three cases:

No custom tiling or pre-routing (ASIC case) With custom tiling only With both custom tiling and pre-routing

Note: only 14% of cells tiled Note: only 3% of wires pre-routed Tiling alone helped reduce the total cell area and eliminate shorts, opens and DRC However, tiling alone increased the total wire count and length

Router can not utilize off-grid routing tracks in sub-array Ends up routing around sub-array

Pre-routing critical nets through the non-standard pitch open tracks was needed

15

Semi-Custom Examples / Pure Data-Path example

Pure Data-Path block tiling example:

32 bit data path consisting of 6 stages of 2:1 muxs Block was implemented first as pure ASIC Block was rebuilt using custom tiling to align each data path bit in straight line In ASIC case BlastFusion optimized the circuit using different types of cells

ASIC case tiling and routing

Semi-Custom case tiling and routing

16

Semi-Custom Examples / Pure Data-Path example (cont.)

no tiling Total cell count Total cell area Utilization Total wire count Total wire length Cells tiled Wires pre-routed Short/Open/DRC
842 0.001763 mm2 89.3% 6558 0.00930 m 0 (0%) 0 (0%) 0/0/0

with tiling
247 0.001161 mm2 58.5% 2650 0.00523 m 196 (77%) 0 (0%) 0/0/0

% smaller
70.6% 34.1% 34.5% 59.6% 43.7% -

Table shows the two cases:

Full ASIC and Tiled data-path

Note: cell count went down by 70% as result of custom tiling ASIC case has high utilization which cases worse placement and routing Tiling reduced utilization by 34% Tiling also resulted in cleaner routing No custom pre-routing was needed

17

Full Custom comparison

Another memory example:

16 KB Data Cache 8 x 2 KB sub arrays Several stages of data path logic Originally built as full custom on previous project Rebuilt as Semi-Custom with identical area and performance Time from RTL to Routed design 1 week

Full Custom

Semi-Custom

18

Full Custom comparison (cont.)

Fully routed Semi-Custom

Metal 4 routes

Metal 5 routes
19

Conclusion

Semi-Custom flow is a powerful design tool:

Improves custom circuit designers productivity Allows control over custom placement and routing Encourages designers to focus on critical parts of the design Utilizes standard Place and Route tools for non-critical parts of design

20