---
id: wiki-2026-0508-bioenergetics
title: Bioenergetics
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [생체 에너지학, ATP, mitochondria, metabolism, thermodynamics, neuromorphic computing]
duplicate_of: none
source_trust_level: B
confidence_score: 0.85
verification_status: conceptual
tags: [biology, biochemistry, atp, metabolism, mitochondria, thermodynamics, neuromorphic, energy-efficiency]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: biology / biochemistry
  applicable_to: [Neuromorphic Computing, Energy-Efficient AI, Drug Discovery]
---

# Bioenergetics

## 📌 한 줄 통찰
> **"매 생명 = 매 entropy 의 저항"**. 매 thermodynamics 2nd law (entropy ↑) 의 against — 매 energy 의 collect + transform → 매 order. 매 ATP 의 currency. 매 modern AI 의 inspiration: 매 neuromorphic computing 의 energy efficiency.

## 📖 핵심

### 매 ATP (생명 의 currency)
- Adenosine Triphosphate.
- 매 phosphate bond 의 cleave → 매 energy.
- 매 cell 의 매 활동 의 fuel.
- 매 인간 의 매 day 의 매 자기 무게 만큼 의 ATP turnover.

### 매 thermodynamics
- **2nd law**: 매 entropy 의 ↑ in 의 closed system.
- **Living system**: 매 open system — 매 외부 의 free energy.
- **Gibbs free energy** (ΔG): 매 work 의 가능 amount.
- **Coupling**: 매 ΔG > 0 reaction 의 매 ΔG < 0 의 hydrolysis 의 drive.

### 매 metabolism

#### Catabolism (이화)
- 매 분해 → 매 energy.
- 매 glucose → 매 36 ATP (full oxidation).

#### Anabolism (동화)
- 매 build → 매 order (protein, DNA).
- 매 ATP 의 consume.

### 매 핵심 pathway
1. **Glycolysis**: 매 cytosol, 매 glucose → pyruvate, 매 2 ATP.
2. **TCA / Krebs cycle**: 매 mitochondria matrix.
3. **Electron transport chain (ETC)**: 매 inner membrane.
4. **Oxidative phosphorylation**: 매 32 ATP.
5. **Fermentation** (anaerobic): 매 lactate / ethanol.

### 매 mitochondria
- 매 powerhouse.
- 매 own DNA (maternal).
- 매 chemiosmotic gradient (Mitchell 1961).
- 매 endosymbiotic origin.

### 매 efficiency
- 매 muscle: 25% (rest 가 heat).
- 매 photosynthesis: 1-3% (광합성).
- 매 brain: 매 20W 의 100T synapse.
- 매 GPU (LLM): 매 100s of W 의 inference.

→ 매 brain 의 efficiency 의 100,000× 의 vs current AI.

### 매 modern AI 의 응용

#### Neuromorphic computing
- 매 spike-based.
- 매 event-driven (sparse).
- 매 in-memory compute.
- 매 chip: Intel Loihi, IBM TrueNorth, BrainChip Akida.

#### Energy-efficient ML
- 매 quantization (INT8, INT4).
- 매 sparse activation.
- 매 mixture of experts (only activated subset).
- 매 distillation.

#### Biological inspiration
- 매 spike-timing-dependent plasticity (STDP).
- 매 reservoir computing.
- 매 differentiable physical system.

### 매 medical 응용
- **Mitochondrial disease**: 매 inherited.
- **Cancer**: 매 Warburg effect (glycolysis 의 prefer).
- **Aging**: 매 mitochondrial dysfunction.
- **Diabetes**: 매 metabolic dysregulation.
- **Drug**: 매 target metabolic enzyme.

### 매 evolutionary
- 매 archaea + bacteria 의 endosymbiosis (mitochondria).
- 매 eukaryote 의 큰 size 의 enable.
- 매 multicellular 의 prerequisite.

## 💻 패턴 (응용 — neuromorphic / energy-efficient ML)

### Spiking Neural Network (SNN)
```python
import torch
import torch.nn as nn

class LIFNeuron(nn.Module):
    """Leaky Integrate-and-Fire — biological-style."""
    def __init__(self, threshold=1.0, decay=0.9):
        super().__init__()
        self.threshold = threshold
        self.decay = decay
        self.v = 0
    
    def forward(self, x):
        self.v = self.decay * self.v + x
        spike = (self.v >= self.threshold).float()
        self.v = self.v * (1 - spike)  # 매 reset on spike
        return spike

# 매 input 의 most 의 zero (sparse) → 매 energy ↓
```

### Energy-aware training
```python
def carbon_aware_train(model, dataset, max_kwh):
    energy_used = 0
    for batch in dataset:
        loss = compute_loss(model, batch)
        loss.backward()
        optimizer.step()
        
        energy_used += measure_gpu_power_wh()
        if energy_used > max_kwh * 1000:
            log(f'Energy budget exhausted: {energy_used} Wh')
            break
```

### Mixture of Experts (sparse activation)
```python
class MoELayer(nn.Module):
    def __init__(self, n_experts=8, top_k=2):
        super().__init__()
        self.experts = nn.ModuleList([Expert() for _ in range(n_experts)])
        self.gate = nn.Linear(d_in, n_experts)
        self.top_k = top_k
    
    def forward(self, x):
        scores = self.gate(x).softmax(dim=-1)
        top_scores, top_idx = scores.topk(self.top_k, dim=-1)
        
        # 매 only top-k 의 active → 매 sparse computation
        out = sum(top_scores[..., i].unsqueeze(-1) * self.experts[idx](x)
                  for i, idx in enumerate(top_idx.unbind(-1)))
        return out
```

### Quantization (INT8 inference)
```python
import torch.quantization as tq

model.eval()
qmodel = tq.quantize_dynamic(model, {nn.Linear}, dtype=torch.qint8)
# 매 75% 의 size ↓, 매 2-4× faster, 매 energy ↓
```

### Energy estimation
```python
# 매 GPU energy of 1 token (LLM)
GPU_TDP_W = 700  # H100
TOKENS_PER_SEC = 1000
ENERGY_PER_TOKEN_J = GPU_TDP_W / TOKENS_PER_SEC  # 0.7 J

# 매 brain comparison
BRAIN_W = 20
BRAIN_TOKEN_EQUIV_PER_SEC = 5  # 매 reading speed
ENERGY_PER_BRAIN_TOKEN_J = BRAIN_W / BRAIN_TOKEN_EQUIV_PER_SEC  # 4 J

# 매 ratio
print(f'GPU 의 brain 의 {ENERGY_PER_BRAIN_TOKEN_J / ENERGY_PER_TOKEN_J:.1f}× efficient per joule')
# 매 surprising X — 매 GPU 가 매 numerical 의 fast 가, 매 brain 의 task 의 different.
```

### Mitochondria simulation (toy)
```python
def atp_yield(glucose, oxygen_present=True):
    """매 simplified glycolysis + TCA + ETC."""
    if oxygen_present:
        glycolysis = 2  # 매 net ATP
        tca = 2 * 1     # 매 GTP
        etc = 2 * 17    # 매 NADH/FADH2 → ATP
        return glycolysis + tca + etc  # 매 ~36
    return 2  # 매 fermentation 만
```

## 🤔 결정 기준 (응용 측)
| 상황 | Approach |
|---|---|
| Edge inference | Quantization + SNN |
| Large model | MoE + sparse |
| Battery-powered | Neuromorphic chip |
| Datacenter | Standard GPU + efficient algorithm |
| Drug discovery | Metabolic pathway model |

**기본값**: 매 sparsity + 매 quantization + 매 hardware 의 right tool.

## 🔗 Graph
- 부모: [[Thermodynamics]]
- 변형: [[ATP]] · [[Mitochondria]] · [[Metabolism]]
- 응용: [[Neuromorphic-Computing]] · [[Mixture-of-Experts]] · [[LLM_Optimization_and_Deployment_Strategies|Quantization]]
- Adjacent: [[Carbon-Footprint]]

## 🤖 LLM 활용
**언제**: 매 energy-efficient AI design. 매 neuromorphic chip exploration. 매 metabolic disease research. 매 carbon-aware ML.
**언제 X**: 매 specific medical claim (의사 consult). 매 nutrition advice.

## ❌ 안티패턴
- **Bigger model only**: 매 energy 의 ignore.
- **Dense everything** (no sparsity): 매 brain 의 inspiration X.
- **Standard FP32**: 매 quantization 의 leverage X.
- **GPU 의 brain 의 mimic 의 expectation**: 매 different paradigm.
- **No carbon tracking**: 매 sustainability ignore.

## 🧪 검증 / 중복
- Verified (Lehninger biochemistry, Mitchell chemiosmotic, Loihi / TrueNorth papers).
- 신뢰도 B.
- Related: [[Neuromorphic-Computing]] · [[Mixture-of-Experts]] · [[LLM_Optimization_and_Deployment_Strategies|Quantization]] · [[Anarcho-Primitivism]] (energy 의 lens).

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — pathway + mitochondria + neuromorphic + 매 SNN / MoE / quantization code |