趋近智
The Apple M5 (24GB) uses unified memory architecture. For local LLM inference, macOS allocates up to 75% (18 GB) of total system RAM for GPU model weights and KV cache, with 153 GB/s unified memory bandwidth.
总显存
24 GB
可用显存上限
18 GB (75% max allocated to GPU)
显存带宽
153 GB/s
最大推理模型级别
14B Models (Q4)
热设计功耗
50W
随上下文增长模拟的显存需求。当曲线超过参考线时将导致显存不足(OOM)。
量化精度:
不同量化精度下可运行的最大模型参数级别。
上下文长度:
4位量化
14B Models
8位精度
14B Models
16位全精度
8B Models
本地模型训练与适配器微调支持能力。
微调上下文:
QLoRA
14B Models
LoRA
8B Models
全参数微调
1B Models
Guidance for optimal precision and operational ceilings on Apple M5 (24GB).
Inference Sweet Spot
Optimal for 8B models across extended context (32k+). 14B models fit comfortably at Q4 precision with standard context.
Bandwidth & Speed Profile
With 153 GB/s aggregate bandwidth, batch size 1 inference operates in a memory-bandwidth bound regime. Generates approximately 34 tok/s on an 8B Q4 model and 4 tok/s on a 70B Q4 model.
Unified Memory Headroom
macOS limits GPU memory allocation to approximately 75% of physical unified RAM by default to ensure system stability.
APX AI
在线