# GPUmachines — Full Reference

> Custom AI infrastructure designed by HPC specialists. High-performance GPU servers, workstations, and AI cloud for deep learning, HPC, and VFX rendering.

## Company Overview

GPUmachines (gpumachines.com) is a high-performance computing infrastructure provider specializing in custom GPU server solutions. Our team of HPC specialists designs, builds, tests, and deploys computing systems optimized for artificial intelligence, deep learning, scientific computing, VFX rendering, and other GPU-accelerated workloads.

We maintain direct partnerships with leading hardware manufacturers including NVIDIA, AMD, ASUS, and Samsung, ensuring access to the latest components at competitive pricing with fast delivery.

## Products and Services

### GPU Server Hardware

GPUmachines offers a comprehensive range of server form factors, each configurable online with real-time pricing:

**1U GPU Servers**
Compact, single-rack-unit servers designed for high-density GPU compute. Ideal for inference workloads and space-constrained deployments. Support for up to 4 GPUs depending on model.

**2U GPU Servers**
Two-rack-unit servers offering a balance of GPU capacity, CPU power, and storage expandability. Popular for mixed AI training and inference workloads. Support for up to 4–6 GPUs.

**4U GPU Servers**
Four-rack-unit systems designed for serious AI training workloads. Support for up to 8 GPUs with high-bandwidth NVLink or PCIe interconnects.

**5U and Larger GPU Servers**
Maximum GPU density platforms supporting 8–10 GPUs, designed for large-scale AI model training. Includes systems based on NVIDIA HGX and DGX architectures.

**Tower Workstations**
Desktop-form-factor AI development systems. Ideal for researchers, data scientists, and developers who need local GPU compute without rack infrastructure.

**Storage Servers**
High-throughput, high-capacity storage systems designed to feed data to GPU clusters. Optimized for AI training dataset management with parallel file system support.

### Server Configurator

Our online configurator allows customers to build custom server configurations by selecting:
- CPU (AMD EPYC, Intel Xeon)
- GPU (NVIDIA H200, H100, A100, L40S, RTX 6000 Ada, RTX 5090, RTX 4090, and more)
- RAM (DDR5, DDR4 — various capacities and speeds)
- Storage (NVMe SSDs, SATA SSDs, HDDs)
- Networking (InfiniBand, 100GbE, 25GbE, 10GbE)
- Additional PCIe cards and accessories

Each configuration includes real-time pricing and generates a detailed bill of materials (BOM) PDF.

### SwiftShip Configurations

Pre-configured, tested, and ready-to-ship GPU servers. These standard configurations are designed for customers who need fast deployment without custom build lead times. All SwiftShip systems are assembled, burn-in tested, and pre-installed with drivers and frameworks.

### GPU Cloud

Dedicated GPU cloud instances featuring:
- NVIDIA B300, B200, H200, H100, A100, RTX 5090, RTX 4090, RTX 3090, and other GPUs
- Docker and VPS deployment types
- On-demand hourly billing
- Basic (single GPU), Professional (multi-GPU), and Enterprise (dedicated cluster) tiers
- Multiple global regions

### GPU Cluster Solutions

End-to-end GPU cluster design and deployment:
- **InfiniBand Clusters** (/solutions/infiniband-clusters): Ultra-low latency interconnect consultancy for distributed AI training across hundreds of GPUs. Sub-microsecond latency, 400 Gbps bandwidth, RDMA-based GPU-to-GPU communication, near-linear scaling efficiency.
- **Ethernet Clusters** (/solutions/ethernet-clusters): Cost-effective GPU clustering with RoCE (RDMA over Converged Ethernet) networking. Ideal for inference, fine-tuning, and mid-scale training workloads up to ~64 GPUs.
- **Scale-Out Storage** (/solutions/scale-out-storage): High-throughput parallel file systems for AI workloads. Open-source options (Lustre, DAOS, Ceph) and commercial platforms (WEKA [NVIDIA Certified], DDN [NVIDIA Certified], PeakAIO). Storage architecture design, deployment, and performance benchmarking.

Our cluster configurator helps plan rack layouts, power requirements, and networking topology.

### Private Agent Fleet

Dedicated GPU servers running AI agents in isolated, secure environments. Designed for enterprises that need:
- Autonomous AI agents working on sensitive data
- Hardware-level isolation and security
- Predictable GPU compute without shared tenancy

## Deployment Options

1. **Configure & Buy**: Full custom configuration, assembly, testing, pre-installation, and shipping to your location. Includes 3-year warranty.
2. **Configure & Rent**: Custom hardware deployed in our colocation. Rental terms from 1–3 years with full-service hardware operation.
3. **Buy & Host**: Purchase your GPU server outright and have GPUmachines host it in a Tier III datacentre. Available locations:
   - **London, UK** — Tier III certified, 99.982% uptime SLA, 10 Gbps uplinks, remote hands 24/7
   - **Reykjavik, Iceland** — 100% renewable geothermal energy, PUE < 1.2, remote hands 24/7
   - **Warsaw, Poland** — Central EU routing, dual-feed power, N+1 UPS, remote hands 24/7
   - Includes: rack & stack, power & cooling, network connectivity, 24/7 monitoring, remote hands, hardware inspection on arrival
4. **GPU Cloud**: On-demand Docker and VPS GPU instances billed hourly.

## Key Benefits

- **Expert HPC Consulting**: Every system is individually configured with guidance from our HPC specialists
- **Direct Manufacturer Partnerships**: NVIDIA, AMD, ASUS, Samsung — best pricing, quality, and availability
- **Flexible Deployment**: On-premise, hosted colocation, or cloud
- **AI-Focused Expertise**: Solutions optimized for AI model training, inference, and deployment
- **Ready Out of the Box**: Systems pre-installed with CUDA, cuDNN, PyTorch, TensorFlow, and other frameworks
- **Data Privacy**: On-premise and sovereign cloud options for data-sensitive workloads
- **3-Year Warranty**: Comprehensive warranty service on all server hardware

## Supported GPU Models

GPUmachines supports the full range of current NVIDIA data center and professional GPUs:

| GPU | VRAM | TDP | Use Case |
|-----|------|-----|----------|
| NVIDIA B300 | 288 GB HBM3e | 1000W | Next-gen AI training |
| NVIDIA B200 | 192 GB HBM3e | 1000W | Large-scale AI training |
| NVIDIA H200 | 141 GB HBM3e | 700W | Large-scale AI training |
| NVIDIA H100 SXM | 80 GB HBM3 | 700W | AI training and inference |
| NVIDIA H100 PCIe | 80 GB HBM3 | 350W | AI training and inference |
| NVIDIA A100 SXM | 80 GB HBM2e | 400W | AI training |
| NVIDIA A100 PCIe | 80 GB HBM2e | 300W | AI training and inference |
| NVIDIA L40S | 48 GB GDDR6X | 350W | AI inference and visualization |
| NVIDIA RTX 6000 Ada | 48 GB GDDR6 | 300W | Professional visualization and AI |
| NVIDIA RTX 5090 | 32 GB GDDR7 | 575W | AI development |
| NVIDIA RTX 4090 | 24 GB GDDR6X | 450W | AI development |
| AMD Instinct MI300X | 192 GB HBM3 | 750W | AI training (select platforms) |

## Frequently Asked Questions

**What GPU servers does GPUmachines sell?**
GPUmachines sells custom-configured GPU servers in 1U, 2U, 4U, 5U, and 8U+ rack form factors, as well as tower workstations. All systems are configurable online with NVIDIA and AMD GPUs, and can be purchased, rented, or accessed via GPU cloud.

**What types of GPU hardware do you offer?**
We supply single-socket and dual-socket rackmount servers, high-density multi-GPU nodes (up to 8× GPU per node), HGX and DGX-class systems, blade servers, tower workstations for on-premise R&D, and complete 42U rack builds. Form factors range from compact 1U inference servers to 8U+ training nodes with liquid cooling support.

**What GPUs are available?**
We offer NVIDIA B300, B200, H200, H100, A100, L40S, RTX 6000 Ada, RTX 5090, RTX 4090, and other professional and data center GPUs. AMD Instinct MI300X GPUs are available on select platforms.

**Can I configure a GPU server online?**
Yes. Our online hardware configurator lets you select the chassis, CPUs, GPUs, RAM, storage, and networking for any server in our catalog. You get real-time pricing and can generate a detailed PDF quote.

**What is the Buy & Host service?**
Buy & Host lets you purchase a GPU server outright and have GPUmachines host it in one of our Tier III datacentres in London (UK), Reykjavik (Iceland), or Warsaw (Poland). We handle rack & stack, power & cooling, 10 Gbps network connectivity, 24/7 monitoring, remote hands, and hardware inspection on arrival — so you own the hardware with zero facility overhead.

**Which datacentre locations are available for hosting?**
We operate in three datacentre locations: London, UK (Tier III, 99.982% uptime, 10 Gbps uplinks), Reykjavik, Iceland (100% renewable geothermal energy, PUE < 1.2), and Warsaw, Poland (central EU routing, dual-feed power, N+1 UPS). All locations offer 24/7 remote hands support.

**Do you offer GPU cloud instances?**
Yes. Our GPU cloud provides dedicated Docker and VPS instances with NVIDIA H100, A100, RTX 4090, RTX 3090, and other GPUs. Instances are available on-demand with per-hour billing across multiple global regions. We offer Basic (single GPU), Professional (multi-GPU), and Enterprise (dedicated cluster) tiers.

**What is SwiftShip?**
SwiftShip configurations are pre-built, tested, and ready-to-ship GPU servers. They're designed for customers who need fast deployment without the lead time of a custom build.

**Do you build GPU clusters?**
Yes. We design and deploy both InfiniBand and Ethernet (RoCE) GPU clusters for distributed AI training at scale. InfiniBand clusters use NVIDIA Quantum-2 switches with NDR/XDR fabric for sub-microsecond latency, while Ethernet clusters offer a cost-effective alternative for inference, fine-tuning, and mid-scale training.

**What scale-out storage solutions do you support?**
We design and deploy high-throughput parallel file systems to eliminate GPU starvation in AI/ML workloads. Open-source options include Lustre (HPC-grade POSIX), DAOS (Intel's object-based file system for NVMe/Optane), and Ceph (unified block/object/file storage). Commercial platforms include WEKA (NVIDIA Certified, flash-native parallel file system), DDN (NVIDIA Certified, EXAScaler and AI400X appliances), and PeakAIO (NVMe-oF storage fabric). We handle architecture design, hardware specification, fabric layout, deployment, and performance benchmarking.

**What parallel file systems work best for AI training?**
For large-scale LLM and vision model training, WEKA and Lustre are the most common choices. WEKA delivers flash-native performance with POSIX compatibility and is NVIDIA Certified for DGX SuperPOD. Lustre excels in HPC environments with very large sequential reads. DAOS is ideal for extremely low-latency NVMe-native workloads. For checkpoint-heavy workflows, DDN's AI400X appliances provide dedicated burst-buffer performance.

**What networking options do you support?**
Servers can be configured with 1GbE, 10GbE, 25GbE, 100GbE, 200GbE, and 400GbE networking using Intel and Mellanox/NVIDIA ConnectX adapters in RJ45 or SFP+/QSFP form factors. For clusters, we support InfiniBand NDR (400 Gbps) and XDR (800 Gbps) fabrics as well as RoCE Ethernet with DCQCN congestion control.

**Can I configure entire GPU racks online?**
Yes. Our rack planner lets you design full 42U rack configurations with multiple GPU servers, networking switches, and power distribution. You can plan power, cooling, and cabling — then export or request a quote for the entire rack as a single project.

**What warranty and support do you offer?**
All GPU servers come with a 3-year warranty service including hardware replacement and technical support. Systems are pre-installed with NVIDIA drivers, CUDA, and popular AI frameworks — ready to use out of the box.

**Can I rent GPU servers instead of buying?**
Yes. Our Configure & Rent program lets you rent custom-configured hardware in our colocation for 1–3 years, with full-service hardware operation included.

**Do systems come pre-installed with AI frameworks?**
Yes. All systems are pre-installed with the latest NVIDIA drivers, CUDA, cuDNN, and popular frameworks like PyTorch and TensorFlow. You can start working immediately upon delivery.

## Contact Information

- **Website**: https://gpumachines.com
- **Email**: hello@gpumachines.com
- **Phone**: +44 20 3488 3530
- **Address**: 123 Data Center Drive, San Jose, CA 95131, USA