AI向けMac Mini M4 vs NVIDIA GPU: ベンチマーク・コスト・徹底比較

# Quick comparison: running Llama 3 8B on each platform

# Mac Mini M4 (Metal via Ollama)
ollama run llama3:8b
# Token generation: ~35 tok/s, Power: ~12W, Cost: $85/mo

# NVIDIA RTX 4090 (CUDA via vLLM)
python -m vllm.entrypoints.openai.api_server \
  --model meta-llama/Meta-Llama-3-8B-Instruct \
  --dtype float16
# Token generation: ~120 tok/s, Power: ~350W, Cost: $500+/mo

# NVIDIA A100 80GB (CUDA via TensorRT-LLM)
trtllm-build --model_dir llama3-8b --output_dir engine
# Token generation: ~180 tok/s, Power: ~250W, Cost: $2,500+/mo

# Reproduce these benchmarks yourself:

# On Mac Mini M4 (using llama-bench)
cd llama.cpp/build
./bin/llama-bench \
  -m ../models/llama-3-8b.Q4_K_M.gguf \
  -ngl 99 -t 8 -p 512 -n 256 -r 5

# Output:
# model                | size   | params | backend | ngl | t/s
# llama-3-8b Q4_K_M    | 4.58 GB| 8.03 B | Metal   | 99  | 35.2 +/- 1.1

# On NVIDIA (using vLLM benchmark)
python benchmark_serving.py \
  --model meta-llama/Meta-Llama-3-8B-Instruct \
  --num-prompts 100 --request-rate 1

# Running Stable Diffusion on Mac Mini M4 with MLX

# Install the MLX Stable Diffusion package
pip install mlx-sd

# Generate an image with SDXL
mlx_sd.generate \
  --model stabilityai/stable-diffusion-xl-base-1.0 \
  --prompt "A futuristic data center powered by renewable energy, photorealistic" \
  --negative-prompt "blurry, low quality" \
  --steps 30 \
  --width 1024 --height 1024 \
  --output generated_image.png

# Batch generation (useful for overnight content pipelines)
for i in $(seq 1 100); do
  mlx_sd.generate --model sdxl-base \
    --prompt "Product photo of a sleek laptop, studio lighting" \
    --output "batch_${i}.png" --seed $i
done

# Run Whisper on Mac Mini M4 using whisper.cpp

# Clone and build whisper.cpp with Metal support
git clone https://github.com/ggerganov/whisper.cpp.git
cd whisper.cpp && make

# Download Whisper Large v3 model
bash ./models/download-ggml-model.sh large-v3

# Transcribe audio (Metal GPU acceleration is automatic)
./main -m models/ggml-large-v3.bin \
  -f meeting-recording.wav \
  --output-txt --output-srt \
  --language en \
  --threads 8

# Result: 10 minutes of audio transcribed in ~45 seconds
# Output: meeting-recording.txt, meeting-recording.srt

# Cost per 1K tokens calculation (Llama 3 8B, 24/7 operation)

# Mac Mini M4 (16GB) - $85/mo
# Throughput: 35 tok/s = 2,100 tok/min = 90.7M tok/mo
# Cost per 1K tokens: $85 / 90,720 = $0.00094
# That's $0.83 per million tokens

# Mac Mini M4 Pro (48GB) - $179/mo
# Throughput: 52 tok/s = 3,120 tok/min = 134.8M tok/mo
# Cost per 1K tokens: $179 / 134,784 = $0.00133
# That's $1.33 per million tokens

# RTX 4090 Cloud - $550/mo
# Throughput: 120 tok/s = 7,200 tok/min = 311.0M tok/mo
# Cost per 1K tokens: $550 / 311,040 = $0.00177
# That's $1.77 per million tokens

# A100 80GB Cloud - $2,535/mo
# Throughput: 180 tok/s = 10,800 tok/min = 466.6M tok/mo
# Cost per 1K tokens: $2,535 / 466,560 = $0.00543
# That's $5.43 per million tokens

# For comparison, OpenAI GPT-4o API:
# Input: $2.50 per million tokens
# Output: $10.00 per million tokens

# Example: nginx load balancer for Mac Mini M4 fleet

upstream llm_backend {
    # Mac Mini M4 fleet (8B models) - always on
    server mac-mini-1.internal:11434 weight=1;
    server mac-mini-2.internal:11434 weight=1;
    server mac-mini-3.internal:11434 weight=1;
}

upstream llm_large {
    # Mac Mini M4 Pro (70B model) - quality tier
    server mac-mini-pro.internal:11434;
}

server {
    listen 443 ssl;
    server_name ai.company.com;

    # Route based on model size header
    location /v1/chat/completions {
        # Default: route to M4 fleet (fast, cheap)
        proxy_pass http://llm_backend;

        # If client requests large model, route to M4 Pro
        if ($http_x_model_tier = "large") {
            proxy_pass http://llm_large;
        }
    }
}

# Monthly cost: 3x M4 ($225) + 1x M4 Pro ($179) = $404/mo
# Equivalent GPU setup: 1x A100 ($2,535) = 6.3x more expensive