NVIDIA
A10 24GB
Tier 1-9
- Memory
- 141 GB HBM3e
- Memory bandwidth
- 4.8 TB/s
- TDP
- Up to 600 W (configurable)
- Interface
- PCIe Gen5 x16, dual-slot
NVIDIA H200 NVL 141GB
P/N900-21010-0040-000
In stockStock tier 10+
Request a quote now
Price and availability confirmed within one business day
- Manufacturer channel supply
- Ships from Dubai / Hong Kong
- Commercial invoice, packing list, warranty docs
01Overview
H200 NVL is the PCIe version of the NVIDIA Hopper data center GPU with 141 GB of HBM3e memory — the largest memory of any PCIe accelerator card. It is bought for LLM inference and fine-tuning of large models, and for HPC workloads in air-cooled servers where SXM-based HGX systems do not fit. Up to four cards can be joined with NVLink bridges to share memory at 900 GB/s.
Built for
- Inference of large language models (70B+ parameters in FP8 on a single card, larger on NVLink-bridged pairs)
- Fine-tuning and retraining of foundation models in an existing PCIe server
- HPC and scientific computing that needs FP64 Tensor Core throughput
- Multi-tenant GPU pools using MIG (up to 7 isolated instances per card)
02Specifications
Technical specifications
Compute
- Architecture
- NVIDIA Hopper
- FP64 Tensor Core
- 60 TFLOPS
- FP32
- 60 TFLOPS
- TF32 Tensor Core
- 835 TFLOPSwith sparsity
- BF16 / FP16 Tensor Core
- 1,671 TFLOPSwith sparsity
- FP8 / INT8 Tensor Core
- 3,341 TFLOPS / TOPSwith sparsity
- GPU clocks
- Base 1,230 MHz, boost 1,785 MHz
- Media engines
- 7 NVDEC, 7 JPEG decoders
Memory
- Capacity
- 141 GB HBM3e
- Bus width
- 6016-bit
- Peak bandwidth
- 4,813 GB/s
- MIG
- Up to 7 instances @ 16.5 GB each
Interface & form factor
- Host interface
- PCIe Gen5 x16 (Gen5 x8 and Gen4 x16 supported), 128 GB/s
- GPU-to-GPU
- 2- or 4-way NVLink bridge, 900 GB/s per GPU
- Form factor
- Full-height, full-length (FHFL) 10.5", dual-slot
- Board weight
- 1,217 g (without bracket, extender and bridge)
- Display outputs
- None (compute card)
Power & cooling
- Total board power
- 600 W maximum (default), 350 W compliance limit, 200 W minimum
- Power connector
- 1× PCIe 16-pin (12VHPWR) auxiliary connector
- Thermal solution
- Passive heat sink — requires server chassis airflow
Software & features
- NVIDIA AI Enterprise
- 5-year subscription included with H200 NVL
- Virtualization
- SR-IOV (32 VF); NVIDIA vGPU 18.1+ (Virtual Compute Server)
- Secure boot
- Supported (CEC1736 root of trust)
- Driver / CUDA
- R565 TRD1 or later; CUDA 12.7 or later
03Requirements
Installation
- Server qualified for H200 NVL (NVIDIA-Certified or MGX partner system) with a free PCIe Gen5 x16 dual-slot FHFL bay
- PCIe 16-pin (12VHPWR) power feed rated for 600 W per card; power mode is configurable down to 350 W
- Passive cooling — chassis fans must deliver airflow for a 600 W card; not suitable for desktops or tower workstations
- NVLink bridges (2-way or 4-way) are ordered separately if GPU-to-GPU memory sharing is needed
In the box
Manufacturer retail or bulk packaging as supplied through the channel; contents confirmed on the order.
04Documents
Manufacturer
Warranty: Manufacturer warranty, details on quote
Supplied with every shipment
- Commercial invoice and packing list
- Manufacturer warranty registration documents
- Certificate of origin on request
- Serial numbers listed on the delivery note
05Related
More in GPUs & accelerators
NVIDIA
A100 80GB SXM
Tier 1-9
Tier 100+
Tier 1-9