Racks and equipment of a compute and network infrastructure
Pepro/GPU servers and AI integration

Companies · laboratories · R&D teams

Your GPU server.
Your AI, in action.

Pepro supplies and prepares your GPU workstations and servers. From choosing the hardware to the first deployed models, we bring together the infrastructure and skills to move your projects forward.

  • Inference & private AI
  • Fine-tuning
  • Shared computing

One contact for hardware, software and commissioning.

Let’s talk on 09 88 06 54 06

A custom configuration

Your use drives the choices.

Hardware and integration

From the GPU to the software environment.

The location of your choice

Your premises or a third-party host.

Based in Grigny

Support in France depending on the assignment; hardware shipped abroad on quotation.

Starting from what you want to do

One infrastructure.
Several ways forward.

Putting a model to work for your team, experimenting or sharing compute: we prepare the environment around your need.

01 / DEPLOY

Your AI working for the team.

A private assistant, document analysis or an application connected to an LLM: we assess the model, its access and the number of requests to handle.

To define: response time, documents, users and expected availability.

02 / EXPERIMENT

From idea to evaluated model.

Fine-tuning, model comparison or quantisation: get an environment to run your experiments, measure results and plan next steps.

To define: data, memory, job duration and quality criteria.

03 / SHARE

Compute for your R&D projects.

Several people, several jobs: we organise resources, access and storage for the compatible computing software chosen.

To define: job queues, priorities, data and application constraints.

Detail of the cooling and housing of graphics cardsHardware · memory · environment

Choosing the GPU goes with the rest

Useful power,
is the one you can actually use.

A powerful card is not enough to make a good server. Memory, processor, storage, network and cooling must work together, with the right tools.

Our role: connecting these choices and delivering an environment prepared for the agreed uses. You keep a clear view of what is supplied, configured and checked.

The four questions to ask before buying →

Before investing

Which server for your workload?

We start with the questions that really change the configuration. You do not need to have chosen a card before contacting us.

Discover our sizing service →
01

Does your model fit in memory?

GPU memory, or VRAM, holds the model’s weights and the data needed to run it. The chosen precision, context length and concurrent requests change the requirement.

Pepro reviews the model, its use and the necessary headroom before settling on the configuration.

02

How many people at the same time?

A company of twenty staff does not necessarily have twenty active requests at once. Usage peaks, the size of exchanges and the expected response time are what help size the service.

We define a representative scenario, then the useful checks at acceptance.

03

One card, several GPUs, or a cluster?

Adding cards does not guarantee your software will use them. Splitting the model or jobs depends on tools, interconnects and compatibility. Memory capacities do not automatically add up into a single usable space.

We choose the architecture from the applications, not just the number of GPUs.

04

Who will run the machine?

Power, cooling, updates, monitoring and intervention: the server’s life is part of the decision. Buying, renting resources or having your machine hosted should be compared over a defined use and duration.

The quote separates hardware, preparation and any recurring services.

Pepro support

The hardware.
The software.
And the link between them.

We can supply the machine and prepare it, or work on your existing hardware. You deal with one contact who connects the equipment to what your team needs to do with it.

A professional looking at a laptop in front of server racks

Illustration. The scope of work is defined with your team.

↗

Size and supply

GPU, CPU, memory, drives and network: a configuration consistent with your applications, budget and planned location. If you already own hardware, we review what it allows.

↗

Prepare a compatible environment

Linux, drivers, CUDA or ROCm depending on the hardware, then Docker and the chosen dependencies. Versions are chosen and checked together.

↗

Put the tools to work

vLLM to serve models, Slurm to organise shared jobs, or the tools useful for your case. Access, users and resources are configured within the agreed scope.

↗

Check and hand over

Representative tests, documented configuration and onboarding: your team knows what has been prepared, how to use it and what remains to monitor. Operational support can extend the assignment.

LinuxCUDA / ROCmDockervLLMSlurmStorage & networking
Close-up of racks and their active equipment

Your infrastructure, your deployment location

At your premises.
Or at your host.

Close to your team

For a workstation or server on site, we review the available environment: location, power, cooling, network and administration.

Prepared before it ships

Is your server going to a third-party host? Pepro can supply it, prepare Linux and the tools, then hand over the information needed to set it up. Form factor, power, connectivity and remote access are checked against the host’s constraints.

Clear responsibilities

Supply, preparation, transport, hosting and follow-up are identified separately. You know who handles each step.

Cooling and power: points to review before deploying (in French) →

Before delivery

What we check.
What you get.

The acceptance scenario is defined with you: application, model, data and conditions of use. Results relate to your configuration, not a universal performance promise.

Criteria for choosing a GPU infrastructure →

The chosen configuration

Hardware references, memory, storage and network; versions of Linux, drivers, CUDA or ROCm and compatible tools. Choices are recorded.

Your application workload

Model, precision or quantisation, exchange length and concurrent requests: test conditions are described along with the results.

Useful measurements

Memory use, response time, time to first token and throughput depending on the case. We check behaviour on the agreed scenario.

Recovery and access

Start-up after reboot, data access and planned permissions; restore procedure if backup is part of the scope.

Handover to your team

Documented configuration, deployment procedure, onboarding and division of responsibilities. Later follow-up is defined separately.

Tests and deliverables are specified in the quote. This page describes our approach; it is not a benchmark already run on your future hardware.

Already equipped? We can start by reviewing your hardware and applications. Storage, networking and software preparation are sized with the rest of the project.

Review my configuration →

Let’s also talk budget

A quote that shows what you are buying.

The card’s price is just one line of the project. We separate investments and services so you can decide on a clear basis.

THE HARDWARE

The configuration

Workstation or server, GPU, memory, storage and network. References, quantities and any additional equipment are specified.

COMMISSIONING

The integration work

Operating system, drivers, tools, access, tests and documentation. The preparation scope is described, with useful options.

WHAT COMES NEXT

Operations

Any hosting, licences, maintenance and support: recurring costs and responsibilities are identified where they apply.

Choose Pepro

A team that connects infrastructure, use cases and handover.

Our work on LLM quantisation addresses a concrete issue: reducing memory and hardware requirements while preserving quality as much as possible. We can extend the integration with a RAG document assistant, from the fine-tuning or training for your team.

Technical training and AI

Auctoritas Formation

Pepro supported Auctoritas Formation with technical training, creating learning content and fine-tuning work.

Discover this project →

Business IT

GFA Pro

GFA Pro entrusts us with managing its IT and its online shop. A relationship that brings together the company’s tools and its day-to-day business.

See the GFA Pro project →

Let’s prepare the first discussion

Where is your project at?

A few pointers are enough to start. Your choices accompany the quote request; we will then refine the sizing with you.

01 · Your main use case
02 · Your equipment
03 · The planned location

Your choices prepare the discussion with Pepro. You can complete your request before sending it; the configuration and price will be set after studying your needs.

Before you start

The questions people ask before buying.

The project can start with a question, a constraint or a model to get running. We help you put the pieces in the right order.

On site, cloud or hosting: our guide →
What budget should we plan for a GPU server?

It depends on the hardware, the model, the number of concurrent users and the integration work. We separate configuration, software preparation and any services. An open need can start with a sizing study before choosing references.

How much GPU memory does my LLM need?

Model size alone does not answer the question. Precision, context, inference engine and concurrent requests change the requirement. We review these parameters and the tests to run before proposing a configuration. quantisation can be considered, with its quality and compatibility trade-offs.

AMD or NVIDIA: how to choose?

We start from your software, models and constraints. CUDA and ROCm are different environments; GPUs, libraries and versions must be compatible. The choice covers the whole chain and what you will actually be able to run.

Can Pepro prepare a server you have already bought?

Yes, work on your existing hardware can be considered. We review the configuration, its condition and the prerequisites for your use before agreeing the scope: operating system, drivers, tools, access and tests.

Can the machine be hosted elsewhere?

Yes. We can prepare your server for deployment at a third-party host, taking its constraints into account. Hosting, access, transport and intervention are specified separately. This Pepro offer is not a hosting contract.

Does our data stay private?

Where the machine sits is only part of the answer. We define with you the access, data flows, services used and responsibilities. Processing can be local or organised in a hosted environment chosen for the project; the architecture must be reviewed as a whole.

And after commissioning?

Documentation and handover are included within the agreed scope. We can consider operational follow-up, upgrades or training for your teams. Hours, terms and any commitments are specified in the proposal.

Your models. Your constraints.
Let’s build the right foundation.

From Paris and its workshop in Essonne, Pepro prepares your infrastructure and supports you through commissioning.

Tell Pepro about my project →

Illustrative photos: Brett Sayles, Nana Dua, Christina Morillo and panumas nikhomkhai / Pexels. Technical references: vLLM sizing and optimisation, GPU compatibility, Slurm.

Your choice is kept for six months in this browser. You can change it at any time via “Manage cookies” at the bottom of the site.

Loading…
Back to top