Long cold aisle of server racks with active blue status LEDs

GPU Platforms

GPU platforms for training, inference and visualisation.

Peregrine designs, sources and commissions GPU systems for infrastructure you own, and deploys them within Peregrine capacity. Architecture and availability are subject to deployment and capacity availability.

Explore platforms Not sure which platform?

Peregrine Compute is a member of the NVIDIA Inception Program.

Platforms

Six platforms, one accountable partner.

Systems are sourced through established OEM and distribution channels; vendor and configuration are selected per project.

Blackwell Ultra

NVIDIA HGX B300

Large-scale training and high-throughput inference for reasoning models.

Designed for

Large-scale LLM training
High-throughput inference
Reasoning and long-context models
Multi-node clusters

Best suited to

Frontier-scale training programmes
Inference fleets serving very large models
Organisations standardising on Blackwell Ultra

GPUs per system: 8 (HGX baseboard)
Memory per GPU: Approx. 288 GB HBM3e (indicative)
Interconnect: NVLink all-to-all across 8 GPUs
Form factor: HGX 8-GPU server, air or liquid cooled

Build with this platform Request capacity

Blackwell

NVIDIA HGX B200

Balanced Blackwell platform for training and production inference.

Designed for

LLM training and fine-tuning
Production inference
Generative AI
Mixed training / inference clusters

Best suited to

Enterprises and AI companies building production clusters
Research groups training large models
Reserved or dedicated capacity

GPUs per system: 8 (HGX baseboard)
Memory per GPU: Approx. 180 GB HBM3e (indicative)
Interconnect: NVLink all-to-all across 8 GPUs
Form factor: HGX 8-GPU server, air or liquid cooled

Build with this platform Request capacity

NVIDIA GB200-class liquid-cooled rack-scale system with coolant manifolds

Rack-scale Blackwell

NVIDIA GB200 / GB300 NVL72-class

Rack-scale, liquid-cooled systems for very large tightly coupled clusters.

Designed for

Very large tightly coupled training
Trillion-parameter-class inference
Rack-scale NVLink domains

Best suited to

Project-specific, deployment dependent
Facilities with liquid-cooling and high power density
Programmes requiring rack-scale NVLink

Architecture: Grace CPU + Blackwell GPUs
NVLink domain: Up to 72 GPUs per rack (indicative)
Cooling: Direct liquid cooling
Form factor: Integrated rack-scale system

Build with this platform Request capacity

Blackwell · PCIe

NVIDIA RTX PRO 6000 Blackwell Server Edition

Cost-effective per-GPU memory for inference fleets, vision, rendering and digital twins.

Designed for

Inference and fine-tuning
Computer vision
Rendering and visualisation
Digital twins and simulation

Best suited to

Inference fleets where memory per dollar matters
Vision, rendering and visualisation workloads
Teams that do not need NVLink coupling

Interface: PCIe Gen5, dual-slot
Memory per GPU: Approx. 96 GB GDDR7 (indicative)
GPUs per server: Typically 4–8
Form factor: 2U–4U PCIe servers

Build with this platform Request capacity

Hopper

NVIDIA HGX H200

Mature ecosystem with larger memory for inference of large models and training.

Designed for

Inference of large models
LLM training and fine-tuning
Memory-bound workloads

Best suited to

Production inference of 70B+ class models
Teams with established Hopper tooling
Reserved capacity

GPUs per system: 8 (HGX baseboard)
Memory per GPU: Approx. 141 GB HBM3e (indicative)
Interconnect: NVLink across 8 GPUs
Form factor: HGX 8-GPU server, air cooled

Build with this platform Request capacity

Hopper

NVIDIA HGX H100

Proven production platform with the broadest software maturity.

Designed for

Training and fine-tuning
Production inference
Research clusters

Best suited to

Universities and research institutes
Established production workloads
On-demand and reserved capacity

GPUs per system: 8 (HGX baseboard)
Memory per GPU: Approx. 80 GB HBM3 (indicative)
Interconnect: NVLink across 8 GPUs
Form factor: HGX 8-GPU server, air cooled

Build with this platform Request capacity

Choosing a platform

Compare at a glance.

Indicative, based on NVIDIA published specifications. Final configuration determined per project.

	B300	B200	NVL72-class	RTX PRO 6000	H200	H100
Memory per GPU (indicative)	~288 GB HBM3e	~180 GB HBM3e	Blackwell GPUs, rack-scale	~96 GB GDDR7	~141 GB HBM3e	~80 GB HBM3
Interconnect	NVLink HGX	NVLink HGX	NVLink rack-scale	PCIe + optional bridge	NVLink HGX	NVLink HGX
Form factor	8-GPU HGX server	8-GPU HGX server	Liquid-cooled rack	2U–4U PCIe server	8-GPU HGX server	8-GPU HGX server
Primary workloads	Training, reasoning inference	Training, inference	Very large coupled clusters	Inference, CV, rendering, twins	Large-model inference, training	Training, inference
Typical cluster scale	Multi-node	Single to multi-node	Rack to multi-rack	Single node to fleets	Single to multi-node	Single to multi-node

Liquid coolant lines and manifold on the rear of a GPU rack

HGX vs PCIe

HGX vs PCIe: when NVLink coupling matters — and when it doesn't.

HGX (NVLink-coupled) platforms connect eight GPUs with high-bandwidth NVLink so they behave like one large accelerator. This matters for large-model training, tensor and pipeline parallelism, and inference of models that do not fit in a single GPU's memory.

PCIe platforms such as RTX PRO 6000 Blackwell suit workloads that scale across independent GPUs: inference fleets serving many model replicas, computer vision pipelines, rendering, digital twins and fine-tuning of models that fit in one GPU. They typically offer more memory per dollar and flexible server form factors.

Most organisations end up with a mix. The Compute Blueprint maps your workloads to the right platform for each.

Not sure which platform?

Let a Compute Blueprint decide.

A fixed-fee assessment sizes your workloads and recommends the platform, interconnect and cluster scale — before you commit. Typical engagement: approximately 2 weeks.

Request a Compute Blueprint

Build with a platform, or request capacity.

Tell us which platform and workload you have in mind. We will confirm fit, deployment options and availability.

Build with Peregrine Request capacity