M MetaX China Last verified

沐曦 曦云 C500

PCIE In production Released 2024 metax-gen1
BF16
TFLOP/s
128 厂商声称
FP8
TFLOP/s
unsupported
FP4
TFLOP/s
unsupported
Memory
GB
64 厂商声称
Mem BW
GB/s
1800 厂商声称
TDP
W
350 厂商声称

Full specs

Compute

FP4 TFLOPS
unsupported
FP8 TFLOPS
unsupported
BF16 TFLOPS
128
FP16 TFLOPS
128
INT8 TOPS
256

Memory

Capacity
64 GB
Bandwidth
1800 GB/s
Type
HBM2e

Die architecture 🟢 vendor floorplan

CU count
64
HBM stacks
4
Process
7 nm
PCIe
Gen 5 ×16

Scale-Up (intra-node)

Protocol
MetaXLink
Per-link BW
200 GB/s
World size
8
Topology
switched
Switch
—

Scale-Out (inter-node)

Per-card NIC
100 Gbps
Protocol
RoCEv2
NIC
—

Topology

拓扑结构 · Topology
8 卡 scale-up domain
芯片内部 / Die-level architecture
HBM HBM HBM HBM 沐曦 曦云 C500 L2 / shared cache · NoC L1$ / register file (per CU) 64 CUs · darker block = tensor / matrix engine 128 TFLOPS BF16 · 64 GB HBM2e @ 1.8 TB/s · 350 W TDP

🟢 vendor floorplan 64 CUs · 4× HBM · 7 nm

集群拓扑 / Cluster topology · MetaXLink @ 200 GB/s
MetaXLink switch 200 GB/s/link · all-to-all GPU 0 64GB GPU 1 64GB GPU 2 64GB GPU 3 64GB GPU 4 64GB GPU 5 64GB GPU 6 64GB GPU 7 64GB 8 cards · switched topology · scale-out: 100 Gbps/card
Scale-Up · 域内
MetaXLink
200 GB/s · 拓扑: switched
world_size = 8
Scale-Out · 跨域
RoCEv2
100 Gbps/卡 NIC

Which models can it run?

Quick estimates · decode tok/s/card 上界

TP=8 · FP16 · batch=16 · prefill=1024 · decode=256 · 已应用 efficiency 校准

在计算器中调整 →
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 63 内存带宽
GLM-5 Reasoning
zhipu
 32B 52 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

Operator-level fit · per-model bottleneck + upper bound

算子级 fit · operator-level fit (per-token roofline)

基于每个模型 operator_decomposition + 本卡 BF16 128 TFLOPS / 1,800 GB/s 计算 · ridge point ≈ 71 FLOPs/byte

上界 = min(计算屋顶, 内存带宽屋顶) · efficiency 未应用
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 21k
GraphCast scientific graph-message-passing 0.9 💾 内存带宽 3321
AlphaFold 3 scientific pair-bias-attention 2.3 💾 内存带宽 998
GPT-OSS llm matmul 0.7 💾 内存带宽 146
Gemma 4 26B llm matmul 0.7 💾 内存带宽 108
DeepSeek V4 Flash llm matmul 0.8 💾 内存带宽 102
Mistral Small 4 llm matmul 0.6 💾 内存带宽 47
Llama 4 Maverick llm matmul 0.8 💾 内存带宽 46
需要 efficiency 校准 + concurrency 扫描 + TCO 估算 → 在计算器中评估 →

Operator support & optimization headroom

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
+-34 pp
saturated

Near saturation at 134% of roofline. Further gains require workload restructure (disaggregated, speculative, smaller batch).

Communication (collective)
All-to-All 🟢 mature
all-to-all via MetaXLink world_size=8
AllReduce 🟢 mature
MetaXLink ring all-reduce
Attention
Multi-Head Attention 🟡 partial
no production attention engine
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 🔴 gap
no MoE-aware engine
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。

海光 DCU Z100
海光信息
100%
相似度
BF16
96
Mem
64
BW
1.0
燧原 云燧 T21
燧原科技
100%
相似度
BF16
192
Mem
64
BW
1.6
壁仞 BR100
壁仞科技
100%
相似度
BF16
256
Mem
64
BW
2.3
寒武纪 思元 590
寒武纪
100%
相似度
BF16
256
Mem
64
BW
1.2

Software-stack support

Engine Status BF16FP16FP4FP8 E4M3FP8 E5M2INT4 AWQ
HanGuangAI unconfirmed — — — — — —
LMDeploy unconfirmed — — — — — —
MindIE unconfirmed — — — — — —
MoRI unconfirmed — — — — — —
SGLang unconfirmed — — — — — —
TensorRT-LLM (Dynamo) unconfirmed — — — — — —
vLLM unconfirmed — — — — — —
Measured efficiency factor

Computed from 1 measured cases for this card. The calculator uses this value in place of the default 0.5.

1.34
measured / theoretical (n=1)

Existing deployment cases (1)

Citations

  1. [1] MetaX 曦云 C500 launch announcement (limited public detail) — https://www.metax-tech.com/ · accessed 2026-04-28 厂商声称
  2. [2] C500 (曦云C500): MACA architecture, 64 compute units, 4× HBM2e ⇒ 64 GB; TSMC 7nm-class — https://www.metax-tech.com/ · accessed 2026-04-28 社区估算
⚠ All performance figures are vendor-claimed unless tier=measured.
⚠ Some fields are best-available estimates from announcements.