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 GPU That Doubles H100 Memory Without Changing the Rest of Your Cluster
The H100 has an 80 GB HBM3 ceiling. The NVIDIA H200 removes it — 141 GB of HBM3e with 4.8 TB/s memory bandwidth, ある 75 percent capacity increase and 60 percent bandwidth increase over H100, in the same SXM5 form factor with the same NVLink topology. For LLM teams whose models exceed 80 GB per GPU — 70B parameter models in full precision, MoE models with large expert weights, or long-context inference with huge KV caches — the H200 is the drop-in upgrade that doubles usable memory per GPU without redesigning the cluster fabric.
技術仕様
| パラメータ | 仕様 |
|---|---|
| GPU Architecture | NVIDIA Hopper (GH100, same die as H100 SXM5) |
| CUDA Cores | 16,896 |
| Tensor Cores | 528 (4第 3 世代, Hopper) |
| メモリー | 141 GB HBM3e |
| Memory Bandwidth | 4.8 TB/s (vs H100’s 3.35 TB/s, +43 percent) |
| Memory Capacity | 141 GB (vs H100’s 80 GB, +76 percent) |
| インタフェース | PCIe 5.0 x16 (SXM5 module via NVSwitch or PCIe variant) |
| NVLink | 900 GB/s (NVSwitch, まで 8 GPUs in one node) |
| FP8 Tensor Core (with sparsity) | 3,958 TFLOPS |
| FP16/BF16 Tensor Core (with sparsity) | 1,979 TFLOPS |
| TF32 Tensor Core | 989 TFLOPS |
| FP64 | 67 TFLOPS |
| Transformer Engine | はい (FP8, automatic precision selection) |
| MIG (Multi-Instance GPU) | まで 7 instances |
| TDP | 700W (SXM5), ~600W (PCIe variant) |
| フォームファクタ | SXM5 module (8-GPU NVLink node) またはPCIe 5.0 dual-slot |
| Supported Platforms | Dell XE9680 (8x H200), xFusion G5500 V7 (up to 8-10x), NVIDIA DGX H200 |
| 保証 | 3-年 (サーバー付き) |
H200 vs H100 vs A100 — The Memory Ceiling Comparison
| 特徴 | H200 141GB | H100 80GB | A100 80GB |
|---|---|---|---|
| メモリー | 141 GB HBM3e | 80 GB HBM3 | 80 GB HBM2e |
| Memory Bandwidth | 4.8 TB/s | 3.35 TB/s | 2.0 TB/s |
| FP8 Tensor (sparsity) | 3,958 TFLOPS | 3,958 TFLOPS (same compute die) | 該当なし (no FP8) |
| 70B Model Fit (FP16, no quant) | はい (141 GB holds ~140 GB weights+KV) | いいえ (needs 2 GPUs or quantization) | いいえ (needs quantization) |
| Long-Context Inference (100K+ tokens) | Excellent (large KV cache fits) | KV cache overflows to CPU | KV cache overflows to CPU |
| Price (relative to H100) | ~135-145 percent | 100 percent | ~50 percent |
Why the 141 GB Matters: The Model Fitting Problem
Training or serving a 70B parameter model in BF16 requires roughly 140 GB of GPU memory — 70B x 2 bytes per parameter. The H100’s 80 GB cannot hold it: you either quantize (losing precision), use tensor parallelism across 2 GPU (doubling memory traffic and halving scaling efficiency), or stream weights (killing throughput). The H200’s 141 GB fits the full model plus KV cache on a single GPU. For inference, a 32K-token context window with a 70B model consumes ~20 GB of KV cache on top of weights — the H200 is the only single-GPU option that fits both without spilling. This is the difference between an inference server that serves one request at a time and one that serves 8-16 concurrent requests.
互換性のあるプラットフォーム
SXM5 variant: Dell XE9680 (8x H200, NVLink NVSwitch fabric), xFusion G5500 V7, NVIDIA DGX H200. PCIe variant: any server with 700W GPU power delivery and adequate airflow. Verify server firmware supports H200 before ordering — H200 requires BIOS/BMC updates on platforms originally shipping with H100.
新川経由のソース
We configure H200 GPU servers with factory-direct pricing, 3-年保証, そして世界的な発送. Lead times are shorter than 2025 H200 allocation; contact us for current availability.
