Nvidia H200 SXM (141GB) AI and HPC GPU

The H200 is not a new architecture. It is the same Hopper GH100 die running the same CUDA cores, the same Tensor Cores, and the same Transformer Engine as the H100. What NVIDIA changed is the memory subsystem: HBM3 replaced with HBM3e, capacity increased from 80GB to 141GB, bandwidth increased from 3,350 GB/s to 4,800 GB/s. Every other specification remains identical. This is a targeted memory upgrade for workloads where the H100's 80GB became the constraint. The practical impact falls into three categories that map to real deployment decisions. LLM inference at full precision. A 70B parameter model at FP16 with KV cache overhead requires approximately 140GB to 160GB of GPU memory depending on batch size and sequence length. The H100 at 80GB cannot fit this without quantization to FP8 or INT8, which reduces model accuracy. The H200 SXM at 141GB fits 70B at FP16 on a single GPU, preserving full precision. For applications where quantization accuracy loss is unacceptable (medical AI, legal document analysis, financial modeling), this is the difference between "possible" and "production-ready." Long-context inference scaling. Transformer models allocate KV cache memory proportional to context window length. A model serving 128K token context windows on the H100 may exhaust 80GB before the sequence completes. The H200's 141GB extends the maximum context window a single GPU can handle by 76 percent before memory offloading becomes necessary. For RAG pipelines, document processing, and conversational agents with long conversation histories, this translates directly to longer contexts without infrastructure complexity. Multi-tenant density via larger MIG partitions. Each MIG instance on the H200 gets 16.5GB versus 10GB on the H100. That 65 percent per-instance increase means each partition can serve larger inference models. Where the H100's 10GB MIG slices handle 3B to 7B models, the H200's 16.5GB slices handle 7B to 13B models per partition. Seven simultaneous 7B model instances on a single H200 GPU is a practical multi-tenant inference configuration. The SXM PCIe form factor details. Up to 4 GPUs connected via NVLink bridges in a single server. Air-cooling compatible at up to 600W TDP configurable (versus 700W on H200 SXM which requires liquid cooling). PCIe Gen 5 x16 interface. Passive heatsink requiring server chassis airflow. Fits the same server platforms running A100 PCIe or H100 PCIe cards, making it a drop-in upgrade. Lenovo, Supermicro, Dell, and HPE all document H200 SXM compatibility on existing server lines. Bandwidth matters for token generation speed. LLM inference token generation in autoregressive decoding is memory-bandwidth-bound: each new token requires reading the entire model weights from HBM. At 4,800 GB/s versus the H100's 3,350 GB/s, the H200 generates tokens approximately 43 percent faster on bandwidth-bound workloads. RunPod benchmarks confirm this translates to real-world inference throughput gains of 1.5x to 1.9x on large language models. The H200 SXM versus H200 NVL decision mirrors the H100 lineup. SXM: 700W TDP, liquid cooling, HGX baseboard, NVSwitch fabric connecting 8 GPUs at 900 GB/s each, optimized for distributed training. SXM: up to 600W, air cooling, standard PCIe servers, up to 4-GPU NVLink, optimized for inference and flexible deployment. SXM for training clusters. SXM for inference servers and existing infrastructure upgrades. Same export controls as the H100. Subject to US restrictions on advanced AI hardware.