Skip to content

Home / Computing / Dedicated GPU servers

Dedicated GPU servers

Choose the GPUs.
Keep the whole server.

A whole server, dedicated to your team. Explore 8-GPU reference configurations from NVIDIA and AMD, then build the right specification for your work.

PLENTYFIELD / DEDICATED
Studio rendering of a closed, graphite server enclosure with machined edges and front service bays
ALLOCATIONA whole server
REFERENCE CONFIGURATIONEight GPUs
SOFTWARE ECOSYSTEMCUDA or ROCm

Compare configurations

Find the right fit.

The right server starts with your software. Compare GPU memory and connectivity, then define the complete build with us.

Choose another GPU

8-GPU DEDICATED SERVER

RTX PRO 6000 Blackwell Server Edition

Inference, rendering, visualization and video. A versatile configuration for work that fits one or a few GPUs.

96 GBMEMORY PER GPU
768 GBTOTAL GPU MEMORY

Eight separate GPU memories; total is not a shared pool.

GPU connection
PCIe Gen 5; no NVLink
Partitioning
Up to four MIG instances per GPU
CPU & RAM
CPU model and system memory specified in your proposal
Local storage
Boot and workload storage specified in your proposal
Networking
2 × 25 GbE front-end Ethernet
Software
Ubuntu LTS · NVIDIA driver · CUDA
Configuration detail

Reference GPU configuration. Host resources and usable storage are specified with the complete server. Connected deployments require a separately designed network.

Request this configuration

Priced in your proposal.

Choose another GPU

8-GPU DEDICATED SERVER

NVIDIA H100 SXM

Production inference, fine-tuning and established GPU workloads with CUDA software support.

80 GBMEMORY PER GPU
640 GBTOTAL GPU MEMORY

Eight separate GPU memories; total is not a shared pool.

GPU connection
NVLink · 900 GB/s per GPU, bidirectional
Memory bandwidth
3.35 TB/s per GPU
CPU & RAM
Selected server platform, sized in your proposal
Local storage
Boot and data storage specified in your proposal
Networking
Front-end Ethernet; dedicated fabric for clusters
Software
Ubuntu LTS · NVIDIA driver · CUDA
Configuration detail

Whole-server reference configuration. Host processors, RAM, storage and delivery depend on the selected platform.

Request this configuration

Priced in your proposal.

Choose another GPU

8-GPU DEDICATED SERVER

NVIDIA H200 SXM

More memory for larger models, longer context and memory-intensive inference.

141 GBMEMORY PER GPU
1,128 GBTOTAL GPU MEMORY

Eight separate GPU memories; total is not a shared pool.

GPU connection
NVLink · 900 GB/s per GPU, bidirectional
Memory bandwidth
4.8 TB/s per GPU
CPU & RAM
Selected server platform, sized in your proposal
Local storage
Mirrored boot and workload storage as specified
Networking
Front-end Ethernet; dedicated fabric for clusters
Software
Ubuntu LTS · NVIDIA driver · CUDA
Configuration detail

Whole-server reference configuration. Final host processors and memory population are validated together.

Request this configuration

Priced in your proposal.

Choose another GPU

8-GPU DEDICATED SERVER

NVIDIA B200

Blackwell computing for large-model inference, training and demanding multi-GPU work.

180 GBMEMORY PER GPU
1,440 GBTOTAL GPU MEMORY

Eight separate GPU memories; total is not a shared pool.

GPU connection
NVLink 5 with NVLink Switch
GPU link bandwidth
1.8 TB/s per GPU, bidirectional
CPU & RAM
OEM platform configuration in your proposal
Local storage
NVMe boot and workload storage
Networking
Front-end Ethernet; fabric for cluster orders
Software
Ubuntu LTS · NVIDIA driver · CUDA
Configuration detail

Whole-server reference configuration. The complete OEM build and service scope are confirmed in your proposal.

Request this configuration

Priced in your proposal.

Choose another GPU

8-GPU DEDICATED SERVER

NVIDIA B300 Blackwell Ultra

High-memory Blackwell Ultra systems for large models and long-context reasoning.

270 GB*MEMORY PER GPU
2,160 GB*TOTAL GPU MEMORY

Eight separate GPU memories; total is not a shared pool.

GPU connection
NVLink 5 with NVSwitch
GPU link bandwidth
1.8 TB/s per GPU, bidirectional
CPU & RAM
Selected OEM platform; confirmed as one build
Local storage
NVMe configuration matched to your workload
Cluster networking
ConnectX-8, up to 800 Gb/s per port; design-specific
Software
Ubuntu LTS · NVIDIA driver · CUDA
Configuration detail

*Memory shown for the Dell 270 GB-per-GPU reference platform. B300 memory figures vary by documented platform; the proposal identifies the exact OEM build and guaranteed specification.

Request this configuration

Priced in your proposal.

Choose another GPU

8-GPU DEDICATED SERVER

AMD Instinct MI325X

A high-memory AMD option for inference and training on compatible ROCm software.

256 GBMEMORY PER GPU
2,048 GBTOTAL GPU MEMORY

Eight separate GPU memories; total is not a shared pool.

GPU connection
AMD Infinity Fabric
Memory bandwidth
6 TB/s per GPU
CPU & RAM
Selected OEM platform, specified in the proposal
Local storage
Boot and data storage to the workload
Networking
Ethernet; dedicated fabric for cluster orders
Software
Ubuntu LTS · ROCm
Configuration detail

Application and framework compatibility are checked before an AMD configuration is proposed.

Request this configuration

Priced in your proposal.

Choose another GPU

8-GPU DEDICATED SERVER

AMD Instinct MI355X

AMD’s high-memory configuration for demanding inference and training workloads.

288 GBMEMORY PER GPU
2,304 GBTOTAL GPU MEMORY

Eight separate GPU memories; total is not a shared pool.

GPU connection
Fourth-generation Infinity Fabric
Memory bandwidth
8 TB/s per GPU
CPU & RAM
Selected OEM platform, specified in the proposal
Local storage
Boot and data storage to the workload
Cooling
Matched to the selected OEM platform and power level
Software
Ubuntu LTS · ROCm
Configuration detail

Cooling, software support and the complete server build are confirmed together for the proposed system.

Request this configuration

Priced in your proposal.

Compare all configurations

Swipe the comparison table to see every specification.

8-GPU serverMemory / GPUTotal GPU memoryGPU connectionSoftware
RTX PRO 6000 Blackwell Server Edition 96 GBGDDR7 ECC768 GBPCIe Gen 5CUDA
NVIDIA H100 SXM 80 GBHBM3640 GBNVLinkCUDA
NVIDIA H200 SXM 141 GBHBM3e1,128 GBNVLinkCUDA
NVIDIA B200 180 GBHBM3e1,440 GBNVLink 5CUDA
NVIDIA B300 Blackwell Ultra 270 GB*HBM3e reference2,160 GB*NVLink 5CUDA
AMD Instinct MI325X 256 GBHBM3E2,048 GBInfinity FabricROCm
AMD Instinct MI355X 288 GBHBM3E2,304 GBInfinity FabricROCm

Total GPU memory adds eight separate GPU memories; it is not one shared memory pool. *B300 uses the documented Dell reference memory specification. The final platform is identified in your proposal.

Reference deployments

Three starting points. Room for your requirements.

Start with one of these configurations, then shape the system around your software, data and operating requirements.

01 / INFERENCE & RENDERING

One dedicated
RTX PRO server.

8 × RTX PRO 6000 Blackwell Server Edition · 96 GB per GPU

For independent jobs and compatible inference, visualization and rendering workloads. Your proposal defines the host, storage, network and access.

Plan this server

02 / CONNECTED AI WORKLOADS

A 32-GPU
B300 cluster.

4 × eight-GPU nodes · one location · dedicated cluster fabric

For work that spans nodes. Servers, switching, data movement and checkpoint storage form one system, with acceptance tests for the complete deployment.

Plan this cluster

03 / AMD COMPUTING

One dedicated
MI355X server.

8 × AMD Instinct MI355X · 288 GB per GPU

For workloads with a compatible ROCm software stack. Your application and a representative job guide the final configuration.

Explore AMD capacity

Reference configurations for proposal discussions. Location, complete build, service scope, price and delivery are confirmed for each order.

NEED MORE THAN ONE SERVER?

Independent jobs.
Or one connected system.

Separate inference replicas and rendering jobs can use independent servers. Work that exchanges data between servers needs an appropriate cluster fabric.

  1. 01
    Choose a connected cluster

    Agree the node count, network topology and storage as a complete system.

  2. 02
    Keep the software fit

    CUDA and ROCm requirements matter as much as GPU memory.

Explore GPU clusters

Whole-server pricing

Know what you’re buying. Each offer sets out the full configuration, location, service, price and term, with delivery timing and any upfront payment stated clearly.

What the service includes

Need a different configuration?

Ask about another GPU, server count, storage arrangement or network.

Request a proposal