Data Center GPU Buying Guide: H100 vs H200 vs B200 vs L40S

By All Chip Supply · Published Oct 9, 2026

Choosing a data center GPU is mostly a question of matching the card to the workload and to the servers you already run. The NVIDIA H100, H200, B200 and L40S are often compared side by side, but they are built for different jobs and different systems. This guide compares them using the figures NVIDIA publishes on its product pages and datasheets, and lists the questions worth answering before you request a quote.

All Tensor Core figures below are NVIDIA's numbers with sparsity unless marked otherwise. Dense throughput is half the sparse figure. Figures differ between SXM modules, PCIe cards and system configurations.

The short version

• L40S is a PCIe card with 48GB of GDDR6 memory, built for mixed AI inference, graphics, rendering and video. It fits standard PCIe servers.
• H100 is the Hopper-generation GPU for large-scale training and inference, sold as an SXM module for HGX systems and as PCIe cards, including the H100 NVL.
• H200 uses the same Hopper compute as H100 but with 141GB of HBM3e memory and 4.8TB/s of bandwidth, which matters most for large models and memory-bound HPC codes.
• B200 is the Blackwell-generation SXM module used in HGX B200 systems, adding FP4 support and much higher throughput per GPU.

Side-by-side specifications

NVIDIA L40S (Ada Lovelace, PCIe Gen4 x16, dual slot, passive cooling)

• Memory: 48GB GDDR6 with ECC, 864GB/s
• FP32: 91.6 TFLOPS
• FP16 / BF16 Tensor Core: 733 TFLOPS
• FP8 Tensor Core: 1,466 TFLOPS
• Max power: 350W
• NVLink: not supported. Multi-Instance GPU (MIG): not supported
• Display outputs and media engines: 4x DisplayPort 1.4a, 3 NVENC and 3 NVDEC with AV1

NVIDIA H100 SXM (Hopper, SXM module)

• Memory: 80GB, 3.35TB/s
• FP64 / FP64 Tensor Core: 34 / 67 TFLOPS
• FP16 / BF16 Tensor Core: 1,979 TFLOPS
• FP8 Tensor Core: 3,958 TFLOPS
• Max TDP: up to 700W (configurable)
• Interconnect: NVLink 900GB/s, PCIe Gen5 128GB/s
• MIG: up to 7 instances at 10GB each

NVIDIA H100 NVL (Hopper, PCIe dual-slot, air-cooled)

• Memory: 94GB, 3.9TB/s
• FP16 / BF16 Tensor Core: 1,671 TFLOPS
• FP8 Tensor Core: 3,341 TFLOPS
• Max TDP: 350–400W (configurable)
• Interconnect: NVLink bridge 600GB/s, PCIe Gen5 128GB/s

NVIDIA H200 SXM (Hopper, SXM module)

• Memory: 141GB HBM3e, 4.8TB/s
• FP64 / FP64 Tensor Core: 34 / 67 TFLOPS
• FP16 / BF16 Tensor Core: 1,979 TFLOPS
• FP8 Tensor Core: 3,958 TFLOPS
• Max TDP: up to 700W (configurable)
• Interconnect: NVLink 900GB/s, PCIe Gen5 128GB/s
• MIG: up to 7 instances at 18GB each

NVIDIA H200 NVL (Hopper, PCIe dual-slot, air-cooled)

• Memory: 141GB HBM3e, 4.8TB/s
• FP8 Tensor Core: 3,341 TFLOPS
• Max TDP: up to 600W (configurable)
• Interconnect: 2- or 4-way NVLink bridge, 900GB/s per GPU, PCIe Gen5 128GB/s

NVIDIA B200 (Blackwell, SXM module in HGX B200) NVIDIA publishes B200 figures for the 8-GPU HGX B200 board. Divided by eight, that is roughly:

• Memory: about 180GB HBM3e per GPU (1.4TB total)
• FP4 Tensor Core: 18 PFLOPS sparse, 9 PFLOPS dense
• FP8 / FP6 Tensor Core: 9 PFLOPS
• FP16 / BF16 Tensor Core: 4.5 PFLOPS
• FP64 / FP64 Tensor Core: 37 TFLOPS
• NVLink (fifth generation): 1.8TB/s GPU-to-GPU

What the numbers mean in practice

Memory capacity decides what fits. A model's weights, the KV cache for inference and the activations for training all need GPU memory. The jump from 80GB (H100 SXM) to 141GB (H200) to around 180GB (B200) can decide whether a model runs on one GPU or must be split across several. The L40S, with 48GB, suits smaller models, fine-tuning and inference at modest batch sizes.

Bandwidth often matters as much as FLOPS. Large language model inference is frequently limited by how fast weights can be read from memory. The H200 has the same compute figures as the H100 SXM but 4.8TB/s of bandwidth against 3.35TB/s, which is why it is often chosen as an inference upgrade within Hopper systems.

Precision support. All four support FP8 Tensor Core math. B200 adds FP4 and FP6. Using these formats depends on your software stack, so check framework and library support before planning around them.

Double precision for HPC. H100 and H200 SXM list 67 TFLOPS of FP64 Tensor Core throughput. The L40S is aimed at AI and graphics and NVIDIA does not list FP64 Tensor Core figures for it, so it is not the right choice for FP64-heavy simulation.

Form factor, power and the server you have

This is often the deciding factor. SXM modules (H100 SXM, H200 SXM, B200) are not standalone cards: they go into HGX baseboards in servers designed for them, with matching power delivery and cooling. They are usually bought as part of complete systems.

PCIe cards (L40S, H100 NVL, H200 NVL, H100 PCIe) fit qualified PCIe servers, but you still need to check:

• Slot space for a dual-slot card
• Power: up to 350W for L40S and 600W for H200 NVL, plus the right 16-pin power cable
• Airflow: these cards are passively cooled and rely on server fans
• Whether the server vendor has qualified the card in that chassis

Questions to answer before you request a quote

1. Which workloads will run: training, inference, rendering, HPC, or a mix?
2. What is the largest model or dataset that must fit in memory on one GPU?
3. Are you buying modules for an HGX system, cards for existing PCIe servers, or complete servers?
4. How many GPUs, and do they need NVLink between them?
5. What power and cooling are available per rack?
6. What delivery timeline and documentation do you need?

Next steps

You can compare listings for the NVIDIA L40S 48GB, H100 NVL 94GB, H100 SXM5 80GB, H100 80GB PCIe, H200 NVL 141GB, H200 SXM 141GB and B200 180GB SXM in our AI and data center category. Availability and lead times change quickly for these parts, so send the details above through a chip request and we will reply with current pricing and timing.

Specifications are taken from NVIDIA's published product pages and datasheets and may change. Always confirm against the latest NVIDIA documentation for your exact part number.