H Huawei Ascend China Last verified

昇腾 910C

OAM In production Released 2024 ascend-910-gen3
BF16
TFLOP/s
800 厂商声称
FP8
TFLOP/s
unsupported
FP4
TFLOP/s
unsupported
Memory
GB
128 厂商声称
Mem BW
GB/s
3200 厂商声称
TDP
W
700 厂商声称

Full specs

Compute

FP4 TFLOPS
unsupported
FP8 TFLOPS
unsupported
BF16 TFLOPS
800
FP16 TFLOPS
800
INT8 TOPS
1600

Memory

Capacity
128 GB
Bandwidth
3200 GB/s
Type
HBM3

Die architecture 🟢 vendor floorplan

AI Core count
64
L2 cache
192 MB
HBM stacks
4
Process
7 nm

Scale-Up (intra-node)

Protocol
HCCS-v2
Per-link BW
784 GB/s
World size
8
Topology
switched
Switch
huawei-hccs-v2-switch

Scale-Out (inter-node)

Per-card NIC
400 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 昇腾 910C L2 / shared cache · NoC L1$ / register file (per AI Core) 64 AI Cores · darker block = tensor / matrix engine 800 TFLOPS BF16 · 128 GB HBM3 @ 3.2 TB/s · 700 W TDP

🟢 vendor floorplan 64 AI Cores · 4× HBM · 192 MB L2 · 7 nm

集群拓扑 / Cluster topology · HCCS-v2 @ 784 GB/s
huawei-hccs-v2-switch 784 GB/s/link · all-to-all GPU 0 128GB GPU 1 128GB GPU 2 128GB GPU 3 128GB GPU 4 128GB GPU 5 128GB GPU 6 128GB GPU 7 128GB 8 cards · switched topology · scale-out: 400 Gbps/card
Scale-Up · 域内
HCCS-v2
784 GB/s · 拓扑: switched
world_size = 8
Scale-Out · 跨域
RoCEv2
400 Gbps/卡 NIC

Which models can it run?

Quick estimates · decode tok/s/card 上界

TP=8 · BF16 · 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 0 内存带宽
Mistral Small 4
mistral
 22B 0 内存带宽
GLM-5 Reasoning
zhipu
 32B 0 内存带宽
GLM-5.1
zhipu
 32B — 显存不足
Qwen3.6 Plus
alibaba
 35B 0 内存带宽
Kimi K2.6
moonshot
 32B — 显存不足
MiniMax M2.7
minimax
 46B 0 内存带宽

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 800 TFLOPS / 3,200 GB/s 计算 · ridge point ≈ 250 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 💾 内存带宽 131k
GraphCast scientific graph-message-passing 0.9 💾 内存带宽 5904
AlphaFold 3 scientific pair-bias-attention 2.3 💾 内存带宽 1774
GPT-OSS llm matmul 0.7 💾 内存带宽 259
Gemma 4 26B llm matmul 0.7 💾 内存带宽 192
DeepSeek V4 Flash llm matmul 0.8 💾 内存带宽 182
Mistral Small 4 llm matmul 0.6 💾 内存带宽 83
Llama 4 Maverick llm matmul 0.8 💾 内存带宽 82
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
+100 pp
big

Currently reaching 0% of theoretical roofline. Massive kernel-tuning headroom — every +0.05 in efficiency ≈ +10% effective throughput.

Communication (collective)
All-to-All 🟢 mature
all-to-all via HCCS-v2 world_size=8
AllReduce 🟢 mature
HCCS-v2 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。

AMD Instinct MI300A
AMD
99%
相似度
BF16
981
Mem
128
BW
5.3
⇄ 海外 替代
Apple M4 Max Neural Engine
Apple
99%
相似度
BF16
38
Mem
128
BW
0.5
⇄ 海外 替代
NVIDIA H200 SXM 141GB
NVIDIA
99%
相似度
BF16
989
Mem
141
BW
4.8
⇄ 海外 替代
Etched Sohu
Etched
99%
相似度
BF16
1125
Mem
144
BW
5.8
⇄ 海外 替代

Software-stack support

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

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

0.00
measured / theoretical (n=1)

Existing deployment cases (1)

Citations

  1. [1] Huawei Ascend 910C announcement (vendor-claimed; specs partially public) — https://e.huawei.com/en/products/computing/ascend · accessed 2026-04-28 厂商声称
  2. [2] Ascend 910C uses Da Vinci 3.0 cores (64 AI Cores per package, dual-die — partially derived from 910B disclosures); HBM3 stacks 4× 32 GB; reportedly SMIC N+2 / 7nm-class process — https://e.huawei.com/en/products/computing/ascend · accessed 2026-04-28 社区估算
⚠ All performance figures are vendor-claimed unless tier=measured.
⚠ Some specs derived from CloudMatrix 384 announcement materials.