NVIDIA L40S
PCIE 在售 发布于 2023 ada-lovelace
BF16
TFLOP/s
366 厂商声称
FP8
TFLOP/s
1466 厂商声称
FP4
TFLOP/s
不支持
Memory
GB
48 厂商声称
Mem BW
GB/s
864 厂商声称
TDP
W
350 厂商声称
完整规格
算力
FP4 TFLOPS
不支持
FP8 TFLOPS
1466
BF16 TFLOPS
366
FP16 TFLOPS
366
INT8 TOPS
1466
显存
容量
48 GB
带宽
864 GB/s
类型
GDDR6
芯片架构 🟢 vendor floorplan
SM count
142
Tensor cores / SM
4
L2 cache
96 MB
制程
5 nm
Die area
609 mm²
Transistors
76 B
PCIe
Gen 4 ×16
Scale-Up (节点内)
协议
PCIe-Gen4
单链带宽
64 GB/s
World size
8
拓扑
pcie-fabric
交换机
—
Scale-Out (节点间)
单卡出口
200 Gbps
协议
InfiniBand-NDR
NIC
ConnectX-7
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 · PCIe-Gen4 @ 64 GB/s
Scale-Up · 域内
PCIe-Gen4
64 GB/s · 拓扑: pcie-fabric
world_size = 8
Scale-Out · 跨域
InfiniBand-NDR
200 Gbps/卡 NIC
ConnectX-7
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
| 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 | 74 | 内存带宽 |
| Mistral Small 4 mistral | 22B | 34 | 内存带宽 |
| GLM-5 Reasoning zhipu | 32B | 28 | 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 366 TFLOPS / 864 GB/s 计算 · ridge point ≈ 424 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 | 💾 内存带宽 | 35k |
| GraphCast | scientific | graph-message-passing | 0.9 | 💾 内存带宽 | 1594 |
| AlphaFold 3 | scientific | pair-bias-attention | 2.3 | 💾 内存带宽 | 479 |
| GPT-OSS | llm | matmul | 0.7 | 💾 内存带宽 | 70 |
| Gemma 4 26B | llm | matmul | 0.7 | 💾 内存带宽 | 52 |
| DeepSeek V4 Flash | llm | matmul | 0.8 | 💾 内存带宽 | 49 |
| Mistral Small 4 | llm | matmul | 0.6 | 💾 内存带宽 | 22 |
| Llama 4 Maverick | llm | matmul | 0.8 | 💾 内存带宽 | 22 |
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
saturated
Near saturation at 150% of roofline. Further gains require workload restructure (disaggregated, speculative, smaller batch).
Communication (collective)
All-to-All 🟢 mature
all-to-all via PCIe-Gen4 world_size=8
AllReduce 🟢 mature
PCIe-Gen4 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 | 官方 | ✓ | ✓ | — | ✓ | ✓ | ✓ |
| MindIE | 未确认 | — | — | — | — | — | — |
| MoRI | 未确认 | — | — | — | — | — | — |
| SGLang | 官方 | ✓ | ✓ | — | ✓ | ✓ | ✓ |
| TensorRT-LLM (Dynamo) | 官方 | ✓ | ✓ | — | ✓ | ✓ | ✓ |
| vLLM | 官方 | ✓ | ✓ | — | ✓ | ✓ | ✓ |
实测校准 efficiency factor
基于 1 个该硬件的实测案例计算得出, 计算器使用此值替代默认 0.5。
1.50
measured / theoretical (n=1)
已有部署案例 (1)
引证
- [1] NVIDIA L40S Datasheet (Ada Lovelace inference-optimized PCIe card) — https://www.nvidia.com/en-us/data-center/l40s/ · 访问于 2026-04-28 厂商声称
- [2] NVIDIA Ada Lovelace Whitepaper (AD102 die: 142 SMs enabled in L40S, 96 MB L2, 76B transistors, 609 mm² @ TSMC 4N) — https://images.nvidia.com/aem-dam/Solutions/Data-Center/l4/nvidia-ada-gpu-architecture-whitepaper-v2.1.pdf · 访问于 2026-04-28 厂商声称
⚠ L40S uses GDDR6 (not HBM); memory bandwidth is much lower than H100/A100, making it bandwidth-bound for many decode workloads.
⚠ Designed for graphics + inference workloads (gaming derivative); strong FP8/INT8 throughput per dollar.
⚠ All performance figures are vendor-claimed unless tier=measured.