NVIDIA H100 PCIe vs SXM vs H100 NVL for local AI

Current eight-H100 workload: Kimi K2.5 fits 8×H100 80GB by aggregate files, but leaves only 44.822GB before runtime allocations.

Short answer: “NVIDIA H100” is not one interchangeable product. Used and integrator listings may refer to an 80GB passive PCIe card, an 80GB SXM module for HGX/DGX platforms, or the 94GB H100 NVL PCIe product. For local AI, first identify the exact variant; only then compare capacity, cooling, power, software and whole-system cost.

H100 PCIe vs SXM vs H100 NVL

Variant	GPU memory	Published memory bandwidth	Power	Deployment clue
H100 PCIe	80GB	2TB/s in NVIDIA’s original datasheet	350W maximum; some 300W-input configurations operate at 310W	Passive, 10.5-inch full-height dual-slot PCIe card
H100 SXM	80GB	3.35TB/s on NVIDIA’s current product page	Up to 700W, configurable	SXM module for HGX H100 or DGX H100 systems
H100 NVL	94GB	3.9TB/s	350–400W, configurable	Dual-slot air-cooled PCIe product

The larger bandwidth figure does not prove a proportional application-speed difference. Quantization kernels, context, batch size, CPU, PCIe topology, storage and runtime configuration all affect measured results.

What 80GB or 94GB means for model fit

Both tiers can hold model artifacts that exceed 48GB, but the printed capacity is not a fit guarantee. Add the exact weights or model file, KV cache, runtime workspaces, graph buffers and safety headroom. A 70B-class model can have very different memory requirements depending on quantization and context. Use the 70B VRAM planning guide and the LLM VRAM calculator before comparing hardware.

The 14GB difference between an 80GB and 94GB product can matter near a memory boundary, but it should not replace measurement. If your workload fits comfortably inside 80GB, platform, cooling and total cost may dominate; if it does not, 94GB still needs enough headroom for the complete workload.

Cooling, power and host requirements

• Passive H100 PCIe 80GB: the heatsink depends on directed chassis airflow. An open workstation case is not evidence of a validated thermal design.
• H100 SXM: requires the matching server platform, baseboard, cooling, power delivery and firmware. A bare module is not a low-cost route to an ordinary desktop.
• H100 NVL: NVIDIA describes it as dual-slot and air-cooled, but it remains a data-center product. Confirm the exact chassis, power mode and qualified system.
• Power connector: verify the board documentation and system cabling for the exact SKU. Do not infer the pinout from a consumer GPU connector that looks similar.

Software, MIG and multi-GPU caveats

• Compute target: NVIDIA’s MIG support table identifies H100 as Hopper with compute capability 9.0. Confirm the required driver, CUDA, framework and compiled kernels.
• MIG: H100 can be partitioned into Multi-Instance GPU profiles. A process assigned a slice sees that slice’s framebuffer, not automatically the full physical GPU.
• NVLink: a high-speed interconnect does not create a transparent pooled-memory device. Your runtime must explicitly split or parallelize the workload. Read how multi-GPU VRAM works.
• Precision claims: sparse peak Tensor Core figures are not tokens-per-second results and should not be compared without matching precision and sparsity conditions.

Used or secondary-market checklist

• Get clear photos of every label; verify capacity, SKU, PCIe versus SXM, and whether “NVL” is actually part of the product name.
• Ask for current device identity, health and memory-error evidence from the exact unit.
• Verify the server or workstation is qualified for the card, including slot, lanes, firmware, airflow, power and management.
• Confirm whether the offer is a bare accelerator, a paired NVL configuration, or a complete tested system.
• Calculate the full deployment cost and review the seller’s return window before paying for the accelerator.

Use the used-GPU buying checklist for a reusable evidence and arrival-test workflow.

Alternatives to compare

Alternative	Why compare it	Key difference
NVIDIA A100 40GB or 80GB	Older Ampere data-center accelerator	Lower-capacity options and different bandwidth/power generations
RTX PRO 6000 Blackwell 96GB	One 96GB workstation CUDA device	Active workstation design, display role and 600W board power
DGX Spark / ASUS Ascent GX10	Complete 128GB unified-memory NVIDIA system	Shared memory and Arm/Linux platform rather than discrete HBM VRAM
AMD Instinct MI210 64GB	Used HBM accelerator alternative	ROCm software lane and 64GB capacity

Official NVIDIA sources

• NVIDIA H100 product page
• NVIDIA H100 GPU datasheet
• NVIDIA H100 PCIe board specification
• NVIDIA MIG supported-GPU table

Would 141GB change the workload? Compare NVIDIA H100 with H200 for local AI, holding Hopper software generation broadly constant while checking the 80GB/94GB-to-141GB capacity step, bandwidth and platform cost.

\n

Reviewed October 25, 2026. Specifications are sourced from NVIDIA and may vary by SKU, revision and configured power mode. HighVRAM has not independently benchmarked H100. Availability and prices change.

Need the next Hopper capacity tier? Compare H200 SXM and H200 NVL for local AI. Both expose 141GB HBM3e, but the SXM and qualified MGX PCIe deployment paths remain distinct.

Moving beyond H100? Compare 141GB H200 with 180GB B200 by complete workload, HGX platform and software compatibility—not peak specifications alone.

What changes from A100? Read the direct A100 vs H100 local-AI comparison for equal-capacity model fit, software generation, bandwidth, power and platform differences.

\n

Need a display-capable 96GB workstation alternative? Compare NVIDIA H100 with RTX PRO 6000 Blackwell across capacity, bandwidth, cooling, MIG and complete-system requirements.

Evaluating H100 for gpt-oss-120b? The gpt-oss VRAM guide separates OpenAI’s single-H100 statement from context, cache and implementation requirements.

Comparing Hopper with Blackwell? Use the H100 vs B200 local-AI guide for exact 80GB-versus-180GB model-fit boundaries, software checks and platform requirements.

Need the full NVIDIA generation ladder? Compare H100, H200, B200 and B300 across 80GB through conservative 270GB capacity, software and platform boundaries.

Would an active 48GB workstation card fit instead? Compare NVIDIA H100 with RTX 6000 Ada across exact model-fit boundaries, HBM bandwidth, cooling, MIG and complete-system requirements.

Would an older active 48GB workstation card fit? Compare NVIDIA H100 with RTX A6000 across exact model-fit boundaries, CUDA generation, HBM bandwidth, cooling, NVLink/MIG and complete-system requirements.

Would a 48GB Ada server card fit instead? Compare H100 with L40S across exact model-fit boundaries, passive cooling, media/vGPU, MIG and interconnect.

Upgrading from an older 48GB server card? Compare NVIDIA A40 with H100 across exact capacity boundaries, Ampere versus Hopper software, MIG, NVLink and complete-system requirements.

Desktop GPU or Hopper accelerator? Compare H100 80GB/94GB with RTX 5090 32GB across exact model-fit, cooling, power, MIG and whole-system requirements.

DGX Spark alternatives: See when H100 is—and is not—a practical DGX Spark alternative in the constraint-first comparison hub.

DGX Spark vs H100: Compare H100 directly with DGX Spark 128GB unified memory in the exact capacity and platform comparison.

Planning gpt-oss-120b? Compare this H100 path with 64GB, 96GB, unified-memory and multi-GPU alternatives using OpenAI’s published memory boundary.

Is H100 the right 80GB path? Use the Best 80GB GPU for local AI guide to compare Hopper capability with A100 value while keeping PCIe, SXM and NVL identities separate.