新川サーバー | エンタープライズ サーバー ハードウェア サプライヤー
The Conversation That Happens in Every AI Infrastructure Meeting
IT Director: “We need GPUs for inference. Engineering wants H100s.”
Infrastructure Architect: “Do they need 80GB of HBM3 and FP8 Transformer Engine for serving a fine-tuned 7B model?”
IT Director: “…They said H100.”
Infrastructure Architect: “They need an L40S. Half the power budget, 30% of the cost, and it handles their inference workload without breaking a sweat. Let us spend the H100 budget on the training cluster where it actually matters.”
This conversation happens because the GPU market segments cleanly into training GPUs and inference GPUs — and confusing the two is the fastest way to overspend on AI infrastructure. The NVIDIA L40S is the inference workhorse: 48 GB GDDR6 with ECC, PCIe Gen4 x16, and 300W TDP in a dual-slot form factor that fits in standard air-cooled servers.
技術仕様
| パラメータ | 仕様 |
|---|---|
| GPU Architecture | NVIDIA Ada Lovelace |
| CUDA Cores | 18,176 |
| Tensor Cores | 568 (4th Gen) |
| RT Cores | 142 (3rd Gen) |
| メモリー | 48 GB GDDR6 with ECC |
| Memory Bandwidth | 864 GB/s |
| Memory Bus | 384-少し |
| インタフェース | PCIe 4.0 x16 |
| FP32 Performance (non-Tensor) | 91.6 TFLOPS |
| TF32 Tensor Core | 183 TFLOPS (with sparsity: 366 TFLOPS) |
| FP16 Tensor Core | 362 TFLOPS (with sparsity: 733 TFLOPS) |
| INT8 Tensor Core | 733 TOPS (with sparsity: 1,466 TOPS) |
| FP8 Tensor Core | 733 TFLOPS (with sparsity: 1,466 TFLOPS) |
| TDP (マックス) | 300W |
| フォームファクタ | Dual-slot, フルハイト, full-length (FHFL), passive cooling |
| Power Connector | 1x PCIe CEM 16-pin (12VHPER) or 1x 8-pin CPU power |
| NVLink / NVSwitch | Not supported (single GPU inference, no multi-GPU fabric) |
| MIG (Multi-Instance GPU) | Not supported (use A100 or H100 for MIG workloads) |
| vGPU Support | NVIDIA vWS, vPC, vApps — supported for VDI and virtualization |
| ECCメモリ | Full ECC on GDDR6 (not all GDDR6 GPUs offer ECC — the L40S does) |
| Physical Dimensions | 267 バツ 112 んん (dual-slot), 1.35 kg |
| 動作温度 | 0℃～45℃ |
| 保証 | 3-year (manufacturer) |
L40S vs H100 vs A100 — Which GPU for Your Workload?
| ワークロード | L40S | A100 80GB | H100 80GB |
|---|---|---|---|
| LLM inference (7B-13B models, FP16) | Excellent. 48GB fits most 13B models. | Overkill unless running multiple models concurrently. | Overkill. FP8 inference is faster but not needed at this scale. |
| LLM inference (70B+ models) | Struggles. 48GB insufficient for 70B models without quantization. | Good with INT8 quantization. 80GB fits 70B models. | Best. FP8 + Transformer Engine for maximum throughput. |
| Full model training (70B+) | Not recommended. No NVLink, limited memory. | Acceptable for fine-tuning. Full training needs NVLink. | Best. NVSwitch + HBM3 for multi-GPU training. |
| Fine-tuning (LoRA, QLoRA, 7B-13B) | Excellent. 48GB + FP16 is sufficient. | Good but more expensive than needed. | Overkill. Spend the savings on more L40S units. |
| VDI / Virtual Desktops | Excellent. vGPU + 48GB supports 6-12 concurrent users. | Good but expensive per-user. | Not designed for VDI workloads. |
| Rendering / RTX workloads | Excellent. RT Cores + 18K CUDA cores. | No RT cores. Not designed for rendering. | No RT cores. |
| HPC simulation (FP64) | Not recommended. No FP64 Tensor Cores. | 良い. FP64 Tensor Cores enabled. | Limited FP64 compared to A100. |
| Price (relative to H100) | ~25-30% | ~55-65% | 100% (baseline) |
Compatible Server Platforms
The L40S is a PCIe 4.0 x16 dual-slot FHFL card compatible with any server that supports 300W GPU power delivery via 8-pin or 12VHPWR connectors. Confirmed compatible platforms include Dell PowerEdge R760xa, R770 (up to 6x L40S), R7715, xFusion G5500 V7, 2288H V7, and 2488H V7. Requires adequate chassis airflow — the passive cooling design depends on server fan speed profiles. Verify your server’s GPU enablement kit includes the correct power cables before ordering.
Source Through Xincuan
We supply L40S GPUs pre-installed in Dell PowerEdge and xFusion servers, or as standalone upgrades for existing platforms. Factory-direct pricing, 3-年保証, and global shipping. Our AI infrastructure engineers can help you right-size your GPU fleet — training cluster vs inference cluster vs mixed workload — to maximize throughput per dollar.
