Dedicated Instant Custom GPU Cloud VPS Unmetered Domains Pricing Data Centers
Marketplace Build your server
A100 · H100 · NVIDIA GPU servers

Enterprise NVIDIA A100 H100 dedicated servers

Bare metal servers with an NVIDIA A100 (40 or 80 GB) or an H100 80 GB, both as PCIe add-in cards, for AI training, inference and high-performance computing. A machine already racked with its card is handed over in minutes; any other build is fitted or bought in, and dated before you pay.

Uptime SLA Enterprise support
View all configurations

NVIDIA A100 & H100 GPU specifications

Enterprise-grade GPU accelerators engineered for AI training, inference, and scientific computing.

Compare technical specifications to select the optimal configuration for your workload requirements.

NVIDIA A100

The A100 GPU delivers exceptional performance, scalability, and reliability for AI training and inference workloads. Built on Ampere architecture with advanced Tensor Cores for accelerated computing at enterprise scale.

Architecture

Ampere

Video memory

40 GB HBM2 or 80 GB HBM2e

CUDA cores

6,912

Max Bandwidth

1,555 GB/s (40 GB), 1,935 GB/s (80 GB)

Max power

250 W (40 GB), 300 W (80 GB)

Bus interface

PCIe Gen 4 x16

NVENC video encoders

None (5 NVDEC decoders)

NVIDIA H100

Hopper, the generation after Ampere, in its PCIe add-in form: the H100 that fits the servers we build. The SXM5 module with HBM3 mounts on an HGX baseboard, and none of our chassis take one, so its figures do not apply here.

Architecture

Hopper

Video memory

80 GB HBM2e

Max Bandwidth

2,039 GB/s

Max power

350 W

Bus interface

PCIe Gen 5 x16

NVENC video encoders

None (7 NVDEC decoders)

Enterprise AI infrastructure for demanding workloads

NVIDIA A100 and H100 dedicated servers powered by Ampere and Hopper architectures, optimized for large-scale AI training, LLM inference, and scientific computing applications.

Ampere architecture

Built on 7nm process with 54 billion transistors, NVIDIA Ampere architecture delivers breakthrough performance for AI training and HPC workloads.

High-bandwidth memory

Stacked memory on every card we fit: HBM2 at 1,555 GB/s on the 40 GB A100, HBM2e at 1,935 GB/s on the 80 GB A100 and at 2,039 GB/s on the H100, all with ECC.

AI acceleration

Advanced Tensor Cores deliver up to 20x performance improvement over previous generations for deep learning training and inference workloads.

Multi-Instance GPU

Partition each GPU into up to seven isolated instances with dedicated compute, memory, and cache resources for optimal multi-tenant utilization.

NVLink connectivity

Both cards carry an NVLink bridge connector. No bridge is on the order form, so two cards in one server talk over PCIe unless a bridge is quoted; ask before ordering if the job needs one.

Enterprise reliability

Data center-grade GPUs with ECC memory, advanced RAS features, and enterprise support for mission-critical production deployments.

FAQ about NVIDIA A100 H100 GPU servers

Common questions about deploying and managing enterprise NVIDIA A100 H100 GPU-accelerated dedicated servers for AI training, inference, and high-performance computing.

What makes NVIDIA A100 and H100 GPUs suitable for enterprise AI workloads?

Both are data-centre accelerators built for arithmetic on large models: stacked HBM memory with ECC, Tensor Cores the major frameworks target, and Multi-Instance GPU for splitting one card between your own jobs. The A100 is Ampere with third-generation Tensor Cores; the H100 is Hopper with fourth-generation Tensor Cores and the Transformer Engine for FP8 precision. Neither card has a hardware video encoder, so neither is the right choice for transcoding.

How do A100 and H100 GPUs compare in performance and capabilities?

As the PCIe cards we fit, the A100 has 6,912 CUDA cores, 40 GB of HBM2 at 1,555 GB/s or 80 GB of HBM2e at 1,935 GB/s, draws 250 or 300 W and connects at PCIe Gen 4. The H100 has 80 GB of HBM2e at 2,039 GB/s, draws 350 W and connects at PCIe Gen 5. H100 figures quoted elsewhere with HBM3 memory describe the SXM5 module, which needs an HGX baseboard that none of our servers use.

What enterprise connectivity and scalability features are available?

Inside one machine, each card can be partitioned into up to seven isolated Multi-Instance GPU instances, each with its own compute, memory and cache. Both cards also carry an NVLink bridge connector, but no bridge is on the order form: two cards in one server talk over PCIe unless a bridge is quoted, so ask before ordering if the job needs one. Between servers the link is each machine's Ethernet port, and multi-GPU and multi-node builds have their own page.

How long does an A100 or H100 server take to deliver?

A machine already racked with the card is delivered within about 30 minutes of payment verification. When we fit a card we hold in stock, it takes 4 to 24 hours; a card we have to buy in takes five to ten working days, or ten to fifteen when the accelerator is on allocation, and you are told which before you pay.