UPF Repeater Strategies

12 min read | Advanced Level

What You'll Learn

• Understand signal integrity challenges at power domain boundaries
• Define repeater insertion strategies using set_repeater
• Specify repeater placement and distance criteria
• Combine repeaters with isolation and level shifters

Signal Integrity at Power Domain Boundaries

Power domain crossings often involve long routing distances, creating signal integrity challenges.

The Long-Wire Problem

Domain A                  Long Wire (>1mm)              Domain B
┌──────────┐              ──────────────────            ┌──────────┐
│  Driver  ├──────────────────────────────────────────> │ Receiver │
└──────────┘         High resistance + capacitance      └──────────┘

Problems:
- Excessive delay (RC delay ∝ distance²)
- Slow slew rate (voltage transitions take longer)
- Signal droop (voltage drops along wire)
- Crosstalk from adjacent wires

RC Delay Scaling

Wire delay follows quadratic relationship with length:

Delay = R × C = (ρ × L) × (ε × L) = ρ × ε × L²

Where:
- R = Wire resistance (∝ length)
- C = Wire capacitance (∝ length)
- L = Wire length

Result: Doubling wire length → 4× delay increase

This quadratic scaling means long wires quickly become timing bottlenecks.

In advanced process nodes (7nm, 5nm), metal resistivity increases due to grain boundary scattering and barrier thickness, making wire delay even worse. Repeater insertion becomes essential for maintaining timing closure on long nets.

What Are Repeaters?

A repeater is a buffer inserted along a long wire to restore signal strength and reduce delay.

Repeater Functionality

Without Repeaters (Long Wire):
Driver ─────────────────────────────────────> Receiver
       └── 3mm wire → 2ns delay, poor slew ──┘

With Repeaters (Buffered):
Driver ──[Buf]──────[Buf]──────[Buf]─────> Receiver
       └─1mm─┘    └─1mm─┘    └─1mm─┘
Total delay: 600ps (optimal spacing reduces delay 3×)

Each repeater:

• Restores signal to full voltage swing
• Improves slew rate (faster transitions)
• Breaks long wire into shorter segments (linear vs quadratic delay)
• Reduces crosstalk sensitivity

Optimal Repeater Spacing

For minimum delay, repeaters should be spaced at:

L_optimal = √(R_buf / (ρ × ε))

Where:
- R_buf = Buffer input resistance
- ρ = Wire resistivity per unit length
- ε = Wire capacitance per unit length

Typical: 500µm - 2mm spacing depending on metal layer and technology

Use physical design tool recommendations for repeater spacing. Modern tools calculate optimal spacing based on actual wire characteristics (metal layer, width, spacing) and buffer library data. UPF repeater strategies guide where repeaters are allowed, not exact spacing.

When to Use Repeater Strategies

Repeaters are beneficial when:

Syntax: set_repeater

The set_repeater command defines repeater insertion strategies for power domain crossings.

Basic Form

set_repeater strategy_name -domain domain_name \
  -applies_to {outputs|inputs|both} \
  [options]

Key Parameters

• strategy_name - Unique identifier for repeater strategy (required)
• -domain domain_name - Domain where repeaters are applied (required)
• -applies_to {outputs|inputs|both} - Signal direction (required)
• -location {self|parent|fanout|automatic} - Repeater placement domain
• -distance value - Maximum wire length before repeater insertion (optional)
• -elements {instance_list} - Specific signals to buffer (optional)

Usage Variations

1. Basic Repeater Strategy:

set_repeater REP_CPU -domain PD_CPU \
  -applies_to outputs \
  -location parent

2. Repeater with Distance Criterion:

set_repeater REP_LONG -domain PD_MEM \
  -applies_to both \
  -location automatic \
  -distance 1000  # Insert repeater if wire > 1mm

3. Repeater for Specific Signals:

set_repeater REP_CRITICAL -domain PD_CPU \
  -applies_to outputs \
  -location self \
  -elements {cpu_inst/critical_sig[*]}

Repeater Placement Strategies

Repeater placement affects signal integrity and which domain supplies power to repeaters.

Self Placement (Source Domain)

set_repeater REP_SELF -domain PD_CPU \
  -applies_to outputs \
  -location self

PD_CPU (source)              PD_TOP (destination)
┌──────────────────┐         ┌─────────────┐
│  Driver──[Rep]───┼────────>│  Receiver   │
│  (in CPU domain) │         │             │
└──────────────────┘         └─────────────┘

Characteristics:

• Repeaters powered by source domain (PD_CPU)
• If source is power-gated, repeaters turn off with domain
• Requires isolation if source can be powered off

Parent Placement (Destination Domain)

set_repeater REP_PARENT -domain PD_CPU \
  -applies_to outputs \
  -location parent

PD_CPU (source)    PD_TOP (destination)
┌──────────┐       ┌──────────────────────┐
│  Driver──┼──────>│ [Rep]──> Receiver    │
│          │       │ (in TOP domain)      │
└──────────┘       └──────────────────────┘

Characteristics:

• Repeaters powered by destination domain (PD_TOP)
• Repeaters remain powered if destination is always-on
• Better for power-gated source domains

Automatic Placement (Tool Decides)

set_repeater REP_AUTO -domain PD_GPU \
  -applies_to both \
  -location automatic

Physical design tool determines optimal repeater placement based on:

• Wire route and congestion
• Power domain availability
• Timing requirements

Use -location parent (destination domain) for repeaters when the source domain can be power-gated. This ensures repeaters remain powered and functional even when the source is off, preventing X-propagation issues.

Distance-Based Repeater Insertion

The -distance option specifies maximum wire length before repeater insertion.

# Insert repeater if wire exceeds 1mm (1000µm)
set_repeater REP_1MM -domain PD_CPU \
  -applies_to outputs \
  -location automatic \
  -distance 1000

Example: Conditional Insertion

The -distance parameter is a guideline, not a strict rule. Physical design tools use this as input for repeater insertion algorithms, combined with actual routing information and timing constraints. Final repeater count and spacing depend on detailed place-and-route results.

Repeaters with Isolation and Level Shifters

Repeaters often combine with other power management strategies.

Repeater + Isolation

For power-gated domains with long wires:

# Power-gated GPU with long outputs
set_repeater REP_GPU -domain PD_GPU \
  -applies_to outputs \
  -location parent         # In always-on domain

set_isolation ISO_GPU -domain PD_GPU \
  -isolation_supply_set SS_TOP \
  -clamp_value 0 \
  -isolation_signal gpu_iso \
  -isolation_sense high \
  -location parent        # Same as repeaters

PD_GPU (gated)         PD_TOP (always-on)
┌──────────┐           ┌────────────────────────┐
│  Driver──┼──────────>│ [ISO]──[Rep]─> Receiver│
│          │           │ (both in TOP domain)   │
└──────────┘           └────────────────────────┘

Order: Isolation first, then repeater (ISO clamps before repeater buffers).

Repeater + Level Shifter

For multi-voltage domains with long distances:

# CPU (1.0V) to TOP (1.2V) with long wire
set_level_shifter LS_CPU -domain PD_CPU \
  -applies_to outputs \
  -location self

set_repeater REP_CPU -domain PD_CPU \
  -applies_to outputs \
  -location parent         # After level shifter
  -distance 1000

PD_CPU (1.0V)          PD_TOP (1.2V)
┌─────────────┐        ┌──────────────────────┐
│  Driver─[LS]┼───────>│ [Rep]──> Receiver    │
│   1.0V→1.2V │ 1.2V   │ (buffered at 1.2V)   │
└─────────────┘        └──────────────────────┘

Order: Level shifter first, then repeater (buffer at destination voltage).

Repeater Library Cell Mapping

Map repeater strategies to physical buffer cells.

Syntax: map_repeater_cell

map_repeater_cell strategy_name -domain domain_name \
  -lib_cells {cell_list}

Example

# Map repeater strategy to various buffer drive strengths
map_repeater_cell REP_CPU -domain PD_CPU \
  -lib_cells {
    BUF_X1      # Minimum drive
    BUF_X2      # Standard drive
    BUF_X4      # High drive
    BUF_X8      # Maximum drive
  }

Physical design tools select appropriate buffer size based on:

• Downstream capacitance
• Fanout
• Timing slack
• Power budget

Specialized Repeater Cells

Some technologies offer low-delay or low-power buffer variants:

# Critical path: Use fast buffers
map_repeater_cell REP_CRITICAL -domain PD_CPU \
  -lib_cells {
    BUF_FAST_X2
    BUF_FAST_X4
    BUF_FAST_X8
  }

# Non-critical: Use low-power buffers
map_repeater_cell REP_GENERAL -domain PD_CPU \
  -lib_cells {
    BUF_LP_X1    # Low-power variant
    BUF_LP_X2
  }

Complete Repeater Example

# ===================================================================
# Repeater Strategy for Large SoC
# ===================================================================

# Domain structure
create_power_domain PD_TOP -include_scope
create_power_domain PD_CPU -elements {cpu_cluster}
create_power_domain PD_GPU -elements {gpu_subsystem}
create_power_domain PD_MEM -elements {mem_ctrl}

# Supply definitions (TOP=1.2V, CPU=1.0V, GPU=0.9V, MEM=1.0V)
# ... (supplies omitted for brevity)

# CPU Repeaters (CPU at 1.0V, TOP at 1.2V)
# Critical signals: Explicit control
set_repeater REP_CPU_CRITICAL -domain PD_CPU \
  -applies_to outputs \
  -location parent \        # In TOP domain (after level shifter)
  -distance 800 \           # Conservative for critical paths
  -elements {cpu_cluster/critical_if[*]}

# General CPU signals: Automatic with relaxed distance
set_repeater REP_CPU_GENERAL -domain PD_CPU \
  -applies_to outputs \
  -location automatic \
  -distance 1500

# CPU inputs from TOP
set_repeater REP_CPU_IN -domain PD_CPU \
  -applies_to inputs \
  -location parent \
  -distance 1000

# GPU Repeaters (power-gated, long distances)
# Place in parent (TOP) domain so repeaters stay powered
set_repeater REP_GPU -domain PD_GPU \
  -applies_to outputs \
  -location parent \        # Always-on domain
  -distance 1000

set_repeater REP_GPU_IN -domain PD_GPU \
  -applies_to inputs \
  -location parent \
  -distance 1000

# MEM Repeaters (same voltage as CPU: 1.0V)
# No level shifter needed between CPU-MEM
set_repeater REP_MEM -domain PD_MEM \
  -applies_to both \
  -location automatic \
  -distance 1200

# Library Cell Mapping
# Critical paths: Fast buffers
map_repeater_cell REP_CPU_CRITICAL -domain PD_CPU \
  -lib_cells {
    BUF_FAST_X2
    BUF_FAST_X4
    BUF_FAST_X8
  }

# General paths: Standard buffers
map_repeater_cell REP_CPU_GENERAL -domain PD_CPU \
  -lib_cells {BUF_X1 BUF_X2 BUF_X4}

map_repeater_cell REP_CPU_IN -domain PD_CPU \
  -lib_cells {BUF_X1 BUF_X2 BUF_X4}

map_repeater_cell REP_GPU -domain PD_GPU \
  -lib_cells {BUF_X1 BUF_X2 BUF_X4 BUF_X8}

map_repeater_cell REP_GPU_IN -domain PD_GPU \
  -lib_cells {BUF_X1 BUF_X2}

map_repeater_cell REP_MEM -domain PD_MEM \
  -lib_cells {BUF_X1 BUF_X2 BUF_X4}

# ===================================================================
# Strategy Summary:
# - Critical paths: Tight distance threshold (800µm), fast buffers
# - General paths: Relaxed threshold (1000-1500µm), standard buffers
# - Power-gated domains: Repeaters in always-on parent domain
# - Multiple drive strengths for optimization flexibility
# ===================================================================

Common Beginner Mistakes

# PROBLEMATIC: Repeaters in GPU domain (can be powered off)
set_repeater REP_GPU -domain PD_GPU \
  -applies_to outputs \
  -location self    # Repeaters lose power when GPU is off!

# BETTER: Place in always-on domain
set_repeater REP_GPU -domain PD_GPU \
  -applies_to outputs \
  -location parent  # Repeaters stay powered

# TOO CONSERVATIVE: Inserting repeaters too frequently
set_repeater REP_EXCESSIVE -domain PD_CPU \
  -applies_to outputs \
  -distance 200   # 200µm is too short for most technologies

# REASONABLE: Based on technology recommendations
set_repeater REP_OPTIMAL -domain PD_CPU \
  -applies_to outputs \
  -distance 1000  # 1mm is typical for modern processes

# SUBOPTIMAL: Single drive strength
map_repeater_cell REP_CPU -domain PD_CPU \
  -lib_cells {BUF_X1}

# OPTIMAL: Multiple options for different loads
map_repeater_cell REP_CPU -domain PD_CPU \
  -lib_cells {BUF_X1 BUF_X2 BUF_X4 BUF_X8}

# ===================================================================
# Repeater Strategy for 6mm CPU-IO Interface
# ===================================================================

# Assume domains and supplies already defined
# CPU at 1.0V, IO at 1.8V, both always-on

# LEVEL SHIFTERS (handle voltage crossing)
set_level_shifter LS_CPU_TO_IO -domain PD_CPU \
  -applies_to outputs \
  -location self    # Shift before long route

set_level_shifter LS_IO_TO_CPU -domain PD_IO \
  -applies_to outputs \
  -location self

# REPEATER STRATEGIES

# Critical 32-bit data bus: Aggressive repeater insertion
set_repeater REP_DATA_BUS -domain PD_CPU \
  -applies_to outputs \
  -location automatic \     # Let tool optimize based on routing
  -distance 800 \           # Repeater every ~800µm (6mm → ~7-8 repeaters)
  -elements {cpu_inst/data_out[31:0]}

# Control signals: Relaxed repeater insertion
set_repeater REP_CONTROL -domain PD_CPU \
  -applies_to outputs \
  -location automatic \
  -distance 1500 \          # Repeater every ~1.5mm (6mm → ~4 repeaters)
  -elements {cpu_inst/ctrl_*}

# Inputs from IO domain (less critical, long distance)
set_repeater REP_CPU_IN -domain PD_CPU \
  -applies_to inputs \
  -location automatic \
  -distance 1200

# IO domain repeaters (symmetric configuration)
set_repeater REP_IO_DATA -domain PD_IO \
  -applies_to outputs \
  -location automatic \
  -distance 800 \
  -elements {io_inst/data_out[31:0]}

set_repeater REP_IO_CONTROL -domain PD_IO \
  -applies_to outputs \
  -location automatic \
  -distance 1500 \
  -elements {io_inst/ctrl_*}

# LIBRARY CELL MAPPING

# Data bus: Fast, high-drive buffers for critical timing
map_repeater_cell REP_DATA_BUS -domain PD_CPU \
  -lib_cells {
    BUF_FAST_X4     # High drive for long routes
    BUF_FAST_X8     # Very high drive option
  }

map_repeater_cell REP_IO_DATA -domain PD_IO \
  -lib_cells {
    BUF_FAST_X4
    BUF_FAST_X8
  }

# Control signals: Standard buffers (non-critical)
map_repeater_cell REP_CONTROL -domain PD_CPU \
  -lib_cells {BUF_X1 BUF_X2 BUF_X4}

map_repeater_cell REP_IO_CONTROL -domain PD_IO \
  -lib_cells {BUF_X1 BUF_X2 BUF_X4}

# Inputs: Standard buffers
map_repeater_cell REP_CPU_IN -domain PD_CPU \
  -lib_cells {BUF_X1 BUF_X2}

# ===================================================================
# Expected Results:
# - Data bus: ~7-8 fast repeaters (aggressive for <1ns timing)
# - Control: ~4 standard repeaters (relaxed for <5ns timing)
# - Total area: ~300-400 repeater cells
# - Timing: Data bus meets <1ns, control easily meets <5ns
# ===================================================================