We're closed from Wednesday 28 October and reopen on Wednesday 4 November. You can still order online, and we'll pick up orders and messages when we're back.
What are you building for?
Tailored recommendationsPC builds across the site will be filtered to match your chosen use-case.

Machine Learning & Data Science

PyTorch & TensorFlow training workstations, for the work that doesn’t fit in the cloud.

ECC memory, dual-GPU options and the I/O bandwidth to keep a training run fed. Built for ML engineers and data scientists who need reproducible performance and on-prem control.

  • ECC RAM by default on Studio+
  • Up to 192GB VRAM (dual RTX 6000 Pro)
  • Threadripper PRO 8-channel memory
  • 5-year warranty
Configure NowTalk to a Specialist

Recommended configurations

Three calibrated starting points. Every component is configurable in our builder.

ML Engineer

Single-GPU training & experimentation

From £9,664

Starting price

CPU
Intel Core Ultra 9 285K
GPU
NVIDIA RTX 5090 (32GB)
RAM
64GB DDR5-6000
Storage
2TB Gen5 NVMe
Best for
Single-GPU training, prototyping
Configure ML Engineer

Most popular

ML Studio

Professional single-GPU, ECC memory

From £24,656

Starting price

CPU
AMD Threadripper PRO 7965WX
GPU
NVIDIA RTX 6000 Pro Blackwell (96GB)
RAM
128GB DDR5 ECC
Storage
4TB Gen5 NVMe
Best for
Reproducible training, long runs
Enquire about ML Studio

ML Production

Dual-GPU, data-team workstation

From £36,591

Starting price

CPU
AMD Threadripper PRO 7995WX
GPU
2× NVIDIA RTX 6000 Pro Blackwell
RAM
256GB DDR5 ECC
Storage
4TB Gen5 NVMe
Best for
Multi-GPU training, large datasets
Enquire about ML Production

GPU vs workload (training)

Training VRAM scales with batch size, sequence length and precision. Mixed precision (bf16 / fp8) and gradient checkpointing reduce VRAM but increase wall-clock time. Figures below are rough sizing guidance.

GPU vs workload (training)
 WorkloadPractical VRAMRecommended tier
Computer vision (CNN/ViT)
16–24GBML Engineer
Tabular / sklearn / XGBoost
CPU + 16GBML Engineer
Fine-tune 7B LLM (LoRA)
24GBML Engineer
Fine-tune 13B LLM (LoRA)
40GBML Studio
Fine-tune 70B LLM (QLoRA)
80GB+ML Studio / Production
Multi-GPU distributed training
2× 96GBML Production

Why CREATE PCs

ECC RAM by default on Studio+

Single-bit memory errors silently corrupt training runs. Every Studio and Production tier ships with registered ECC DDR5.

I/O that keeps the GPU fed

Gen5 NVMe primary, large secondary NVMe for datasets. Optional 10GbE for shared dataset stores.

Linux-first if you need it

Ubuntu LTS preinstalled with CUDA, cuDNN, conda and your chosen framework. Full warranty either way.

ML Training Workstations FAQs

Workstation or cloud: which is cheaper for training?

A simple rule of thumb: if you train more than ~6 hours a day on a single high-end GPU, on-prem typically pays back inside 12 months versus equivalent cloud pricing. Add the convenience of fast local data access and reproducibility, and most ML teams keep a serious workstation for daily work and burst to cloud for huge runs.

Do you support Linux for training?

Yes, Ubuntu 22.04 / 24.04 LTS is supported with CUDA, cuDNN, conda and PyTorch / TensorFlow preinstalled on request. Warranty is identical to Windows builds.

Why Threadripper PRO over Ryzen 9?

Two reasons: 8-channel ECC memory (vs 2-channel non-ECC on AM5), and far more PCIe lanes (important for dual GPUs, fast NVMe RAID and 10GbE simultaneously). For single-GPU workstations Ryzen 9 9950X is plenty; once you go dual GPU or want ECC, Threadripper PRO is the right answer.

Can I run two RTX 6000 Pro cards in one chassis?

Yes, we build dual-RTX-6000-Pro stations in workstation cases with 1600W Platinum PSUs and validated chassis airflow. Sustained dual-GPU loads are tested for 24 hours before dispatch.

Do you offer 10GbE / Thunderbolt for shared storage?

Yes. We can spec 10GbE on Threadripper PRO builds and Thunderbolt 4 on Intel-based builds for connection to NAS / DAS dataset stores.