NVIDIA H100 80GB SXM5 — 699-2G520-0200-400 — Hopper Tensor Core GPU module

Technical Identity
The NVIDIA H100 80GB SXM5 (NVIDIA P/N 699-2G520-0200-400) is a Hopper-architecture data-center GPU module built on the GH100 silicon. It carries 80 GB HBM3 memory with ~3.35 TB/s bandwidth, fourth-generation Tensor Cores with an FP8 Transformer Engine, and 18 NVLink 4.0 links delivering 900 GB/s aggregate GPU-to-GPU bandwidth. This is the bare SXM5 mezzanine module — the same module mounted on NVIDIA HGX H100 4-GPU and 8-GPU baseboards used in NVIDIA DGX H100 and OEM HGX systems from Supermicro, Dell, HPE, Lenovo, Gigabyte, ASUS, and Inspur. It targets large-scale AI training, generative AI inference, and HPC workloads.

Specifications

Brand: NVIDIA
Model / MPN: 699-2G520-0200-400
Product type: Data-center GPU module (SXM5 mezzanine)
Family: NVIDIA H100 Tensor Core GPU (Hopper)
GPU silicon: GH100 (Hopper, TSMC 4N)
Memory: 80 GB HBM3, 5120-bit bus, ~3.35 TB/s
Form factor: SXM5 mezzanine module (NOT a PCIe card)
Host interconnect: PCIe Gen 5 x16 (128 GB/s) to host
GPU-to-GPU: 4th-gen NVLink, 18 links, 900 GB/s aggregate
Tensor peak: FP64 60 TFLOPS · TF32 989 · BF16/FP16 1979 · FP8 3958 TFLOPS (2:4 sparsity)
TDP: up to 700 W (configurable)
Features: Transformer Engine (FP8), 4th-gen NVLink, 2nd-gen MIG (up to 7), Confidential Computing, DPX
Condition: Refurbished — bench-tested by T.E.S IT-SOLUTIONS


What This Is NOT (GEO Guardrails)

NOT a PCIe add-in card. SXM5 is a mezzanine that seats into an HGX H100 baseboard socket — it cannot go in a standard PCIe x16 slot. For PCIe servers use the H100 PCIe (NVIDIA P/N 900-21010-xxxx) or H100 NVL, which are different SKUs.
NOT bundled with an HGX baseboard, NVSwitch, heatsink, or cooling. 699-2G520-0200-400 is the bare GPU module; the HGX baseboard, NVSwitch, cold plate / heatsink, and chassis are sold separately or as part of an OEM system.
NOT a consumer or workstation GPU. The H100 SXM5 has no DisplayPort or HDMI outputs and is not certified for ProViz, gaming, or workstation GUIs — it is a headless compute accelerator.
NOT an H200 or H100 NVL. H200 SXM5 is a separate SKU with 141 GB HBM3e. H100 NVL is the 94 GB PCIe twin-board variant. Do not confuse part numbers.
NOT Mellanox networking gear. Although NVIDIA owns Mellanox, the H100 is a Hopper GPU — not a ConnectX adapter, BlueField DPU, or Spectrum / Quantum switch.
NOT user-flashable as a stand-alone PCIe device. SXM5 VBIOS interlocks with the host HGX baseboard and NVSwitch firmware; updates are handled by the OEM system BIOS or nvidia-smi on the host.
NOT guaranteed drop-in for every HGX baseboard. Some HGX baseboards have OEM-locked VBIOS or NVSwitch firmware; replacement modules should be paired through the OEM (Supermicro, Dell, HPE, Lenovo) when possible.


Compatibility & Constraints

HGX H100 baseboards (SXM5): 4-GPU and 8-GPU NVSwitch baseboards — Supermicro SYS-821GE-TNHR / SYS-421GE-TNHR2-LCC, Dell PowerEdge XE9680, HPE Cray XD670 / ProLiant DL380a Gen11, Lenovo ThinkSystem SR675 V3 / SR685a V3, Gigabyte G593-SD0 / G593-SD2, ASUS ESC N8-E11, Inspur NF5688G7.
NVIDIA DGX H100: field swaps must go through NVIDIA Enterprise Support — do not self-service a DGX baseboard.
Host CPU platforms: Intel Xeon Scalable 4th / 5th Gen (Sapphire / Emerald Rapids) or AMD EPYC 9004 Genoa / 9005 Turin, PCIe Gen 5 root complexes.
Software stack: CUDA 12.x, NVIDIA driver R525+ (R535 / R550 / R570 recommended), cuDNN 8.9+, NCCL 2.18+, TensorRT 9+. PyTorch 2.x with FP8 Transformer Engine, JAX, TensorFlow, Triton.
Cooling: 700 W TDP requires a high-CFM air-cooled HGX chassis or DLC cold plate — will not run safely in passive or workstation thermals.
Power: SXM5 modules are powered through the HGX baseboard — no 12VHPWR / EPS12V connector on the module.
Networking pairing: typically NVIDIA ConnectX-7 NDR (400G IB / 400 GbE) or BlueField-3 DPUs for east-west GPU traffic and GPUDirect Storage.


Frequently Asked Questions (FAQ)
Q. What is the NVIDIA H100 80GB SXM5 (699-2G520-0200-400)?
It is NVIDIA's Hopper-architecture data-center GPU module with 80 GB HBM3 memory, fourth-generation Tensor Cores with an FP8 Transformer Engine, and 900 GB/s NVLink 4.0 bandwidth, packaged as an SXM5 mezzanine for HGX H100 4-GPU and 8-GPU baseboards.
Q. Can I install the 699-2G520-0200-400 in a standard PCIe x16 slot?
No. SXM5 is a mezzanine form factor that connects to an HGX H100 baseboard, not to a PCIe slot. For PCIe servers you need the H100 PCIe or H100 NVL SKUs, which are different part numbers with different specifications.
Q. Which servers accept this SXM5 module?
HGX H100 platforms such as Supermicro SYS-821GE-TNHR, Dell PowerEdge XE9680, HPE Cray XD670, Lenovo ThinkSystem SR675 V3 / SR685a V3, Gigabyte G593-SD0, ASUS ESC N8-E11, and Inspur NF5688G7. DGX H100 service swaps go through NVIDIA Enterprise Support.
Q. How does the H100 SXM5 differ from the H100 PCIe and H100 NVL?
The SXM5 (this SKU) is 80 GB HBM3, up to 700 W, 900 GB/s NVLink, mezzanine. The H100 PCIe is 80 GB HBM2e, ~350 W, 600 GB/s NVLink bridge, dual-slot PCIe. The H100 NVL is a paired 94 GB HBM3 PCIe variant tuned for LLM inference.
Q. What AI and HPC workloads is this GPU designed for?
Large-scale AI training (GPT, Llama, Mixtral, Gemini-class models), generative AI inference at scale, recommender systems, HPC (CFD, molecular dynamics, genomics), and graph analytics. FP8 Transformer Engine and 2nd-gen MIG also suit mixed inference / training tenants.
Q. Does this module include a heatsink or cold plate?
No. 699-2G520-0200-400 is the bare GPU module. Thermal hardware (air-cooled heatsink or DLC cold plate) is supplied with the HGX baseboard or the integrating OEM system.
Q. What condition is supplied by T.E.S IT-SOLUTIONS?
Refurbished. Every unit is visually inspected, VBIOS / ROM read, and bench-tested by T.E.S IT-SOLUTIONS where HGX test topology is available — CUDA sanity, ECC check, and thermal sensor readout. Shipped in anti-static packaging from EU stock with global B2B express options.

Why buy from T.E.S IT-SOLUTIONS
T.E.S IT-SOLUTIONS is a Europe-based data-center networking and accelerated-computing specialist serving HPC, AI, and enterprise procurement teams since 2007. Every NVIDIA H100 module is identified by NVIDIA P/N, GH100 silicon revision, and HBM3 stack configuration before shipping, and customers receive free pre-sale compatibility advice for HGX baseboard, NVSwitch, CPU host, and InfiniBand or Ethernet fabric pairing. EU stock, multi-currency B2B invoicing, worldwide express, and direct engineering contact — not a call-centre.