Apple M4 Max Neural Engine
EMBEDDED-SOC 在售 发布于 2024 apple-neural-engine-gen5
BF16
TFLOP/s
38 厂商声称
FP8
TFLOP/s
不支持
FP4
TFLOP/s
不支持
Memory
GB
128 厂商声称
Mem BW
GB/s
546 厂商声称
TDP
W
11 厂商声称
完整规格
算力
FP4 TFLOPS
不支持
FP8 TFLOPS
不支持
BF16 TFLOPS
38
FP16 TFLOPS
38
INT8 TOPS
38
显存
容量
128 GB
带宽
546 GB/s
类型
LPDDR5X
芯片架构 🟢 vendor floorplan
NeuralEngine count
16
制程
3 nm
Scale-Up (节点内)
协议
UMA
单链带宽
546 GB/s
World size
1
拓扑
unified
交换机
—
Scale-Out (节点间)
单卡出口
0 Gbps
协议
none
NIC
—
拓扑示意
拓扑结构 · Topology
单卡 / single card
芯片内部 / Die-level architecture
⚠ illustrative / 示意性版图: compute-unit and HBM-stack count are inferred from public BF16 / memory specs. architecture field not populated for this card yet. Contribute floorplan data →
集群拓扑 / Cluster topology · UMA @ 546 GB/s
Scale-Up · 域内
UMA
546 GB/s · 拓扑: unified
world_size = 1
Scale-Out · 跨域
none
0 Gbps/卡 NIC
能跑哪些模型?
Quick estimates · decode tok/s/card 上界
TP=1 · BF16 · batch=16 · prefill=1024 · decode=256 · 已应用 efficiency 校准
| 模型 | 参数 (active) | Decode tok/s/card | 瓶颈 |
|---|---|---|---|
| DeepSeek V4 Pro deepseek | 49B | — | 显存不足 |
| DeepSeek V4 Flash deepseek | 13B | — | 显存不足 |
| Mistral Small 4 mistral | 22B | — | 显存不足 |
| GLM-5 Reasoning zhipu | 32B | 6 | 内存带宽 |
| GLM-5.1 zhipu | 32B | — | 显存不足 |
| Qwen3.6 Plus alibaba | 35B | — | 显存不足 |
| Kimi K2.6 moonshot | 32B | — | 显存不足 |
| MiniMax M2.7 minimax | 46B | — | 显存不足 |
算子级 fit · 任意模型瓶颈类型 + 上界
算子级 fit · operator-level fit (per-token roofline)
基于每个模型 operator_decomposition + 本卡 BF16 38 TFLOPS / 546 GB/s 计算 · ridge point ≈ 70 FLOPs/byte
| 模型 | domain | 主导算子 | AI · F/B | 瓶颈 | tok/s 上界 |
|---|---|---|---|---|---|
| DeepSeek V4 Pro | llm | matmul | 245.5 | 🔥 计算 | 6318 |
| GraphCast | scientific | graph-message-passing | 0.9 | 💾 内存带宽 | 1007 |
| AlphaFold 3 | scientific | pair-bias-attention | 2.3 | 💾 内存带宽 | 303 |
| GPT-OSS | llm | matmul | 0.7 | 💾 内存带宽 | 44 |
| Gemma 4 26B | llm | matmul | 0.7 | 💾 内存带宽 | 33 |
| DeepSeek V4 Flash | llm | matmul | 0.8 | 💾 内存带宽 | 31 |
| Mistral Small 4 | llm | matmul | 0.6 | 💾 内存带宽 | 14 |
| Llama 4 Maverick | llm | matmul | 0.8 | 💾 内存带宽 | 14 |
需要 efficiency 校准 + concurrency 扫描 + TCO 估算 → 在计算器中评估 →
算子支持 & 优化空间
算子支持 & 优化空间 / Operator support & headroom
Per-operator support derived from software_support.engines + scale-up topology. Optimization headroom from measured efficiency factor.
Optimization headroom
+50 pp
moderate
No cases yet — using default 0.5 efficiency. Real headroom unknown until first measurement lands.
Communication (collective)
All-to-All 🟡 partial
small scale-up domain; expert-parallel needs careful sharding
AllReduce 🟢 mature
UMA ring all-reduce
Attention
Multi-Head Attention 🟢 mature
paged-attention via vLLM/SGLang/MindIE
FlashAttention-3 🔴 gap
No FA-3 path; falls back to FA-2 / vanilla SDPA
Matrix multiply (GEMM)
Matrix Multiplication 🟢 mature
GEMM supported on all inference engines
MoE routing
MoE Routing 🟢 mature
MoE gating supported via vLLM ≥0.4 / SGLang
Normalization
RMSNorm 🟢 mature
fused into engine kernels
Embedding
Rotary Position Embedding 🟢 mature
fused into engine kernels
Activation
SiLU / Swish 🟢 mature
fused into engine kernels
Softmax 🟢 mature
fused into engine kernels
最接近的替代卡 (按规格相似度)
基于 BF16 算力 / 显存 / 显存带宽 / FP8 加权欧氏距离。供选型决策参考。
软件栈支持
| 引擎 | 状态 | BF16 | FP16 | FP4 | FP8 E4M3 | FP8 E5M2 | INT4 AWQ |
|---|---|---|---|---|---|---|---|
| HanGuangAI | 未确认 | — | — | — | — | — | — |
| LMDeploy | 未确认 | — | — | — | — | — | — |
| MindIE | 未确认 | — | — | — | — | — | — |
| MoRI | 未确认 | — | — | — | — | — | — |
| SGLang | 未确认 | — | — | — | — | — | — |
| TensorRT-LLM (Dynamo) | 未确认 | — | — | — | — | — | — |
| vLLM | 社区 | ✓ | ✓ | — | — | — | ✓ |
已有部署案例 (0)
暂无该硬件的实测案例。
成为第一个贡献者?
引证
- [1] Apple M4 Max: 16-core Neural Engine @ 38 TOPS, unified memory architecture (UMA) up to 128 GB LPDDR5X with 546 GB/s bandwidth shared with GPU/CPU. ~11 W package TDP for ANE portion. — https://www.apple.com/mac/m4/ · 访问于 2026-04-29 厂商声称
⚠ Apple M4 NPU rating is 38 TOPS at INT8 precision; FP16 is roughly 38 TFLOPS via Metal Performance Shaders.
⚠ Unified memory means the 128 GB capacity is shared across CPU/GPU/ANE; for inference, expect ~80-100 GB usable.
⚠ No multi-card scaling: single-package only. World-size=1 reflects this.
⚠ Energy efficiency (3.4 TOPS/W) is the strongest of any accelerator in the corpus — relevant for edge / on-device.