Etched Sohu
PCIE announced 发布于 2025 etched-sohu-gen1
BF16
TFLOP/s
1125 厂商声称
FP8
TFLOP/s
2250 厂商声称
FP4
TFLOP/s
4500 厂商声称
Memory
GB
144 厂商声称
Mem BW
GB/s
5760 厂商声称
TDP
W
700 厂商声称
完整规格
算力
FP4 TFLOPS
4500
FP8 TFLOPS
2250
BF16 TFLOPS
1125
FP16 TFLOPS
1125
INT8 TOPS
2250
显存
容量
144 GB
带宽
5760 GB/s
类型
HBM3e
芯片架构 🟢 vendor floorplan
Tile count
144
制程
4 nm
PCIe
Gen 5 ×16
Scale-Up (节点内)
协议
Etched-Mesh
单链带宽
800 GB/s
World size
8
拓扑
full-mesh
交换机
—
Scale-Out (节点间)
单卡出口
400 Gbps
协议
Ethernet-RoCE
NIC
—
拓扑示意
拓扑结构 · Topology
8 卡 scale-up domain
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 · Etched-Mesh @ 800 GB/s
Scale-Up · 域内
Etched-Mesh
800 GB/s · 拓扑: full-mesh
world_size = 8
Scale-Out · 跨域
Ethernet-RoCE
400 Gbps/卡 NIC
能跑哪些模型?
Quick estimates · decode tok/s/card 上界
TP=8 · FP4 · 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 | 117,551 | 内存带宽 |
| DeepSeek V4 Flash deepseek | 13B | 164 | 内存带宽 |
| Mistral Small 4 mistral | 22B | 75 | 内存带宽 |
| GLM-5 Reasoning zhipu | 32B | 62 | 内存带宽 |
| GLM-5.1 zhipu | 32B | 42 | 内存带宽 |
| Qwen3.6 Plus alibaba | 35B | 40 | 内存带宽 |
| Kimi K2.6 moonshot | 32B | 34 | 内存带宽 |
| MiniMax M2.7 minimax | 46B | 27 | 内存带宽 |
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 1,125 TFLOPS / 5,760 GB/s 计算 · ridge point ≈ 195 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 | 187k |
| GraphCast | scientific | graph-message-passing | 0.9 | 💾 内存带宽 | 11k |
| AlphaFold 3 | scientific | pair-bias-attention | 2.3 | 💾 内存带宽 | 3193 |
| GPT-OSS | llm | matmul | 0.7 | 💾 内存带宽 | 466 |
| Gemma 4 26B | llm | matmul | 0.7 | 💾 内存带宽 | 346 |
| DeepSeek V4 Flash | llm | matmul | 0.8 | 💾 内存带宽 | 328 |
| Mistral Small 4 | llm | matmul | 0.6 | 💾 内存带宽 | 149 |
| Llama 4 Maverick | llm | matmul | 0.8 | 💾 内存带宽 | 147 |
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 🟢 mature
all-to-all via Etched-Mesh world_size=8
AllReduce 🟢 mature
Etched-Mesh ring all-reduce
Attention
Multi-Head Attention 🟢 mature
paged-attention via vLLM/SGLang/MindIE
FlashAttention-3 🟢 mature
FA-3 on modern engine + tensor cores
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] Etched Sohu (announced June 2024): Transformer-only ASIC, claims 100,000+ tokens/sec on Llama 70B (8-card system), 144 GB HBM3e per chip. Status: announced, GA targeted late 2025. — https://www.etched.com/announcing-etched · 访问于 2026-04-29 厂商声称
- [2] Sohu architecture estimate: ~144 specialized Tiles optimized for transformer attention + MLP only. Cannot run non-transformer workloads (no graph ops, no conv, no MoE gate primitive without firmware extension). — https://www.semianalysis.com/p/sohu-asic-deep-dive · 访问于 2026-04-29 社区估算
⚠ Sohu is TRANSFORMER-ONLY: cannot run scientific (AlphaFold), graph (GraphCast), or vision (SAM/DINO) workloads. Domain restriction is the entire bet.
⚠ Status: announced (June 2024); GA pushed to late 2025 / 2026 — specs subject to revision at GA.
⚠ Vendor-claimed throughput numbers (100k+ tok/s on Llama 70B) imply ~10x H100 efficiency — independent verification pending.