趋近智
The Apple M2 Max (96GB) uses unified memory architecture. For local LLM inference, macOS allocates up to 75% (72 GB) of total system RAM for GPU model weights and KV cache, with 400 GB/s unified memory bandwidth.
总显存
96 GB
可用显存上限
72 GB (75% max allocated to GPU)
显存带宽
400 GB/s
最大推理模型级别
70B+ Models (Q4)
热设计功耗
100W
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
70B+ Models
8位精度
32B Models
16位全精度
32B Models
本地模型训练与适配器微调支持能力。
微调上下文:
QLoRA
70B+ Models
LoRA
70B Models
全参数微调
7B-8B Models
Guidance for optimal precision and operational ceilings on Apple M2 Max (96GB).
Inference Sweet Spot
Suitable for 70B parameter models at Q4/Q8 with extended 32k context windows, or high-throughput batching of smaller dense architectures.
Bandwidth & Speed Profile
With 400 GB/s aggregate bandwidth, batch size 1 inference operates in a memory-bandwidth bound regime. Generates approximately 89 tok/s on an 8B Q4 model and 11 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