S SambaNova Systems Last verified

SambaNova SN40L

RECONFIGURABLE 在售 发布于 2024 sambanova-rdu-gen4
BF16
TFLOP/s
638 厂商声称
FP8
TFLOP/s
1040 厂商声称
FP4
TFLOP/s
不支持
Memory
GB
1664 厂商声称
Mem BW
GB/s
6400 厂商声称
TDP
W
300 厂商声称

完整规格

算力

FP4 TFLOPS
不支持
FP8 TFLOPS
1040
BF16 TFLOPS
638
FP16 TFLOPS
638
INT8 TOPS
1040

显存

容量
1664 GB
带宽
6400 GB/s
类型
HBM3

芯片架构 🟢 vendor floorplan

RDU-Tile count
1040
制程
5 nm
Die area
800 mm²
PCIe
Gen 5 ×16

Scale-Up (节点内)

协议
SambaFabric
单链带宽
800 GB/s
World size
16
拓扑
switched
交换机

Scale-Out (节点间)

单卡出口
200 Gbps
协议
Ethernet
NIC

拓扑示意

拓扑结构 · Topology
16 卡 scale-up domain
芯片内部 / Die-level architecture
HBM HBM HBM HBM HBM HBM HBM HBM SambaNova SN40L L2 / shared cache · NoC L1$ / register file (per RDU-Tile) 1040 RDU-Tiles · darker block = tensor / matrix engine 638 TFLOPS BF16 · 1040 FP8 · 1664 GB HBM3 @ 6.4 TB/s · 300 W TDP

⚠ 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 →

🔄 reconfigurable dataflow (RDU)

集群拓扑 / Cluster topology · SambaFabric @ 800 GB/s
ToR · SambaFabric Node 1 Node 2 2 节点 × 8 卡 = 16 卡 · 节点内 800 GB/s · 节点间 200 Gbps RoCE/IB
Scale-Up · 域内
SambaFabric
800 GB/s · 拓扑: switched
world_size = 16
Scale-Out · 跨域
Ethernet
200 Gbps/卡 NIC

能跑哪些模型?

Quick estimates · decode tok/s/card 上界

TP=8 · FP8 · 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 86,451 计算
DeepSeek V4 Flash
deepseek
13B 182 内存带宽
Mistral Small 4
mistral
22B 83 内存带宽
GLM-5 Reasoning
zhipu
32B 69 内存带宽
GLM-5.1
zhipu
32B 47 内存带宽
Qwen3.6 Plus
alibaba
35B 45 内存带宽
Kimi K2.6
moonshot
32B 38 内存带宽
MiniMax M2.7
minimax
46B 30 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 638 TFLOPS / 6,400 GB/s 计算 · ridge point ≈ 100 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 106k
GraphCast scientific graph-message-passing 0.9 💾 内存带宽 12k
AlphaFold 3 scientific pair-bias-attention 2.3 💾 内存带宽 3548
GPT-OSS llm matmul 0.7 💾 内存带宽 517
Gemma 4 26B llm matmul 0.7 💾 内存带宽 385
DeepSeek V4 Flash llm matmul 0.8 💾 内存带宽 364
Mistral Small 4 llm matmul 0.6 💾 内存带宽 166
Llama 4 Maverick llm matmul 0.8 💾 内存带宽 164
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 SambaFabric world_size=16
AllReduce 🟢 mature
SambaFabric 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
fused into engine kernels
Activation
SiLU / Swish 🟢 mature
fused into engine kernels
Softmax 🟢 mature
fused into engine kernels

软件栈支持

引擎 状态 BF16FP16FP4FP8 E4M3FP8 E5M2INT4 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. [1] SambaNova SN40L (Cardinal): RDU (Reconfigurable Dataflow Unit) with 1040 PCUs, 64 MB on-chip SRAM, 1.5 TB HBM3 + DDR5 hybrid memory (1664 GB total), 5nm TSMC, 800 mm² die. Full system: 8-card SN40L node with 12 TB aggregate fast memory. — https://sambanova.ai/products/sn40l · 访问于 2026-04-29 厂商声称
⚠ SN40L unique architecture: 3-tier memory (on-chip SRAM + HBM + DDR5) lets it host 5+ trillion-parameter models in a single node — no other accelerator does this.
⚠ Compute reported is per-RDU; sustained throughput depends on dataflow graph fit (reconfigurable; not all models map cleanly).