case_study_cooling_decision.py

#!/usr/bin/env python3
"""
Case Study: The Cooling Upgrade That Wouldn't Help.

SCENARIO:
  A data center architecture team is designing the next-generation AI
  accelerator on a 5 nm process at 1.5 GHz. Their current silicon die uses
  standard server air cooling (h = 1,000 W/m²K). Compute density is at the
  thermal limit — they can't pack more gates without overheating.

  The team is evaluating two options:
    Option A: Upgrade to direct liquid cooling (h = 20,000 W/m²K).
              Cost: ~$2M for a 10,000-unit data center retrofit.
    Option B: Stay with air cooling but switch the substrate to SiC.
              Cost: higher per-die cost, but no data center retrofit.

  Management wants to know: which option unlocks more compute density?

NON-OBVIOUS CONCLUSION:
  Option A (expensive liquid cooling) improves maximum density by only 0.3%.
  Option B (SiC substrate with the same cheap air cooling) improves it by 232%.

  The reason is the conduction floor — an irreducible thermal resistance set
  by the substrate's thermal conductivity. On silicon, heat cannot leave the
  die fast enough regardless of how aggressively you cool the surface. No
  amount of convective cooling can overcome this physics.

  Aethermor surfaces this conclusion in under 15 seconds. Without it, the
  team would need to set up and run separate COMSOL models for each material
  and cooling configuration — typically 1-2 days of engineering time.

BONUS INSIGHT:
  Even without changing the substrate or cooling, the team can gain 47% more
  throughput by redistributing compute density from the thermally-bottlenecked
  GPU block to the thermally-underutilized cache and I/O blocks. The L3 cache
  has 26x thermal headroom. The I/O complex has 14x. The GPU is at the wall.

  The power budget (200 W) is not even close to binding — only 10 W is used.
  The constraint is thermal, not electrical. This is invisible without a tool
  that models both power and thermal simultaneously.

RUN IT:
  python benchmarks/case_study_cooling_decision.py
"""

import sys
import os
import time

sys.stdout.reconfigure(encoding="utf-8", errors="replace")

from aethermor.analysis.thermal_optimizer import ThermalOptimizer
from aethermor.physics.chip_floorplan import ChipFloorplan, FunctionalBlock
from aethermor.physics.materials import get_material

# ── Configuration ─────────────────────────────────────────────────────
TECH_NODE = 5       # nm
FREQ = 1.5e9        # Hz
H_AIR = 1000.0      # server air cooling, W/(m²·K)
H_LIQUID = 20000.0  # direct liquid cooling, W/(m²·K)
H_EXOTIC = 50000.0  # exotic direct-die cooling, hypothetical upper bound

opt = ThermalOptimizer(tech_node_nm=TECH_NODE, frequency_Hz=FREQ)

# ══════════════════════════════════════════════════════════════════════
print("=" * 72)
print("CASE STUDY: The Cooling Upgrade That Wouldn't Help")
print("=" * 72)
print()
print(f"  Scenario: {TECH_NODE} nm AI accelerator at {FREQ/1e9:.1f} GHz")
print(f"  Current cooling: server air (h = {H_AIR:.0f} W/m²K)")
print(f"  Question: upgrade to liquid cooling, or change substrate?")
print()


# ── Part 1: Cooling upgrade barely helps on silicon ───────────────────
print("-" * 72)
print("PART 1: How much does better cooling actually buy on silicon?")
print("-" * 72)

t0 = time.perf_counter()
si_air = opt.find_max_density("silicon", h_conv=H_AIR)
si_liquid = opt.find_max_density("silicon", h_conv=H_LIQUID)
si_exotic = opt.find_max_density("silicon", h_conv=H_EXOTIC)
elapsed_1 = time.perf_counter() - t0

gain_liq = (si_liquid["max_density"] - si_air["max_density"]) / si_air["max_density"] * 100
gain_exo = (si_exotic["max_density"] - si_air["max_density"]) / si_air["max_density"] * 100

print(f"\n  Computed in {elapsed_1:.1f}s\n")
print(f"  {'Cooling Solution':<30s}  {'h (W/m²K)':>12s}  {'Max Density':>14s}  {'Gain vs Air':>12s}")
print(f"  {'-'*30}  {'-'*12}  {'-'*14}  {'-'*12}")
print(f"  {'Server air (current)':<30s}  {H_AIR:>12,.0f}  {si_air['max_density']:>14.3e}  {'baseline':>12s}")
print(f"  {'Direct liquid ($2M retrofit)':<30s}  {H_LIQUID:>12,.0f}  {si_liquid['max_density']:>14.3e}  {gain_liq:>11.1f}%")
print(f"  {'Exotic direct-die (theoretical)':<30s}  {H_EXOTIC:>12,.0f}  {si_exotic['max_density']:>14.3e}  {gain_exo:>11.1f}%")

print()
print(f"  ⚠ RESULT: 20× more aggressive cooling buys only {gain_liq:.1f}% more density.")
print(f"  Even a 50× cooling upgrade (exotic direct-die) gains only {gain_exo:.1f}%.")
print()
print(f"  WHY: Silicon's thermal conductivity (148 W/mK) creates an irreducible")
print(f"  conduction floor. Heat cannot leave the die interior fast enough,")
print(f"  regardless of how aggressively the surface is cooled.")


# ── Part 2: Substrate switch is transformative ────────────────────────
print()
print("-" * 72)
print("PART 2: What if we change the substrate instead?")
print("-" * 72)

t0 = time.perf_counter()
materials = ["silicon", "silicon_carbide", "gallium_nitride", "diamond"]
ranking = opt.material_ranking(h_conv=H_AIR, materials=materials)
elapsed_2 = time.perf_counter() - t0

si_d = si_air["max_density"]

print(f"\n  Computed in {elapsed_2:.1f}s  (same server air cooling for all)\n")
print(f"  {'Substrate':<25s}  {'k (W/mK)':>10s}  {'Max Density':>14s}  {'vs Silicon':>12s}  {'vs Si+Liquid':>14s}")
print(f"  {'-'*25}  {'-'*10}  {'-'*14}  {'-'*12}  {'-'*14}")
for r in ranking:
    mat = get_material(r["material"])
    ratio_si = r["max_density"] / si_d
    ratio_liq = r["max_density"] / si_liquid["max_density"]
    marker = ""
    if r["material"] == "silicon":
        marker = " (baseline)"
    print(f"  {r['material_name']:<25s}  {mat.thermal_conductivity:>10.0f}  {r['max_density']:>14.3e}"
          f"  {ratio_si:>11.1f}×  {ratio_liq:>13.1f}×{marker}")

sic = next(r for r in ranking if r["material"] == "silicon_carbide")
sic_gain_pct = (sic["max_density"] - si_d) / si_d * 100
sic_vs_liquid = sic["max_density"] / si_liquid["max_density"]

print()
print(f"  ✓ RESULT: SiC with basic air cooling gives {sic_gain_pct:.0f}% more density than")
print(f"     silicon with the same air cooling — and {sic_vs_liquid:.1f}× more than silicon")
print(f"     with the $2M liquid cooling upgrade.")
print()
print(f"  THE DECISION: Don't retrofit the data center. Change the substrate.")
print(f"  SiC + air ≫ Si + liquid. The physics makes this unambiguous.")


# ── Part 3: Free throughput from compute redistribution ───────────────
print()
print("-" * 72)
print("PART 3: What can we gain for FREE right now? (no hardware changes)")
print("-" * 72)

# Build a representative SoC
fp = ChipFloorplan(
    grid_shape=(40, 40, 6),
    element_size_m=50e-6,
    material="silicon",
)
fp.add_block(FunctionalBlock(
    name="GPU_cluster", x_range=(0, 25), y_range=(0, 25), z_range=(0, 6),
    gate_density=1e4, activity=0.6, tech_node_nm=TECH_NODE, paradigm="cmos",
))
fp.add_block(FunctionalBlock(
    name="L3_cache", x_range=(25, 40), y_range=(0, 25), z_range=(0, 6),
    gate_density=8e3, activity=0.03, tech_node_nm=TECH_NODE, paradigm="cmos",
))
fp.add_block(FunctionalBlock(
    name="IO_complex", x_range=(0, 40), y_range=(25, 40), z_range=(0, 6),
    gate_density=2e3, activity=0.15, tech_node_nm=TECH_NODE + 2, paradigm="cmos",
))

# Headroom analysis
t0 = time.perf_counter()
headroom = opt.thermal_headroom_map(fp, h_conv=H_AIR)
elapsed_3a = time.perf_counter() - t0

print(f"\n  Thermal headroom analysis (computed in {elapsed_3a:.3f}s):\n")
print(f"  {'Block':<20s}  {'T_max (K)':>10s}  {'Headroom':>10s}  {'Status':>15s}")
print(f"  {'-'*20}  {'-'*10}  {'-'*10}  {'-'*15}")
for h in headroom:
    hf = h.get("density_headroom_factor", 0)
    if h.get("is_bottleneck"):
        status = "← BOTTLENECK"
    elif hf > 10:
        status = f"← {hf:.0f}× underused"
    else:
        status = ""
    print(f"  {h['name']:<20s}  {h['T_max_K']:>10.1f}  {hf:>10.1f}×  {status:>15s}")

# Power redistribution
t0 = time.perf_counter()
result = opt.optimize_power_distribution(
    fp, power_budget_W=200.0, frequency_Hz=FREQ, h_conv=H_AIR
)
elapsed_3b = time.perf_counter() - t0

print(f"\n  Power-constrained optimization (computed in {elapsed_3b:.3f}s):\n")
if "optimised_blocks" in result:
    print(f"  {'Block':<20s}  {'Original':>12s}  {'Optimised':>12s}  {'Change':>10s}")
    print(f"  {'-'*20}  {'-'*12}  {'-'*12}  {'-'*10}")
    for br in result["optimised_blocks"]:
        change = br.get("density_change", 1)
        arrow = "↑" if change > 1.01 else ("↓" if change < 0.99 else "=")
        print(f"  {br['name']:<20s}  {br['original_density']:>12.0f}  "
              f"{br['optimised_density']:>12.0f}  {arrow} {change:>8.1f}×")

improvement = result.get("improvement_ratio", 1)
total_power = result.get("total_power_W", 0)
power_budget = result.get("power_budget_W", 200)
binding = result.get("binding_constraint", "unknown")

print()
print(f"  Throughput improvement: {improvement:.2f}× ({(improvement - 1) * 100:.0f}%)")
print(f"  Power used: {total_power:.1f} W of {power_budget:.0f} W budget")
print(f"  Binding constraint: {binding}")
print()
print(f"  ✓ RESULT: {(improvement - 1) * 100:.0f}% more throughput with ZERO hardware changes.")
print(f"  The L3 cache block has massive thermal headroom — the optimizer shifts")
print(f"  compute density there, pulling it from the thermally-limited GPU block.")
print()
print(f"  CRITICAL INSIGHT: The power budget ({power_budget:.0f} W) is not the constraint.")
print(f"  Only {total_power:.1f} W of {power_budget:.0f} W is used — the binding constraint is THERMAL.")
print(f"  Adding more electrical power capacity does nothing. The physics bottleneck")
print(f"  is heat removal from the die interior, not power delivery to it.")


# ── Summary ───────────────────────────────────────────────────────────
total_time = elapsed_1 + elapsed_2 + elapsed_3a + elapsed_3b
print()
print("=" * 72)
print("SUMMARY: Three Non-Obvious Engineering Conclusions")
print("=" * 72)
print(f"""
  1. COOLING UPGRADES DON'T HELP ON SILICON.
     Spending $2M on liquid cooling buys {gain_liq:.1f}% more density.
     The conduction floor makes this nearly zero-return.

  2. SUBSTRATE CHANGE IS 700× MORE EFFECTIVE THAN COOLING CHANGE.
     SiC + cheap air cooling gives {sic_gain_pct:.0f}% gain vs Si + air.
     Si + expensive liquid cooling gives {gain_liq:.1f}% gain vs Si + air.
     Ratio: {sic_gain_pct / max(gain_liq, 0.001):.0f}× more effective.

  3. COMPUTE REDISTRIBUTION GIVES {(improvement - 1) * 100:.0f}% FOR FREE.
     The GPU block is at the thermal wall. The cache block has {headroom[1].get('density_headroom_factor', 0):.0f}×
     headroom. Redistributing density improves throughput {improvement:.2f}×
     without changing the substrate, cooling, or power budget.

  WHAT THE ENGINEER WOULD HAVE DONE WITHOUT THIS DATA:
     Ordered the liquid cooling retrofit. Lost $2M. Gained 0.3%.

  WHAT THIS TOOL SHOWS IN {total_time:.0f} SECONDS:
     Skip the cooling upgrade. Either switch to SiC or just
     redistribute compute density across blocks for immediate gains.

  This is the kind of insight that changes an engineering decision —
  and the reason architecture-stage thermal exploration matters.
""")