Skip to content
← WorkProject 008Concept build

Vanta Workstation

A dual-GPU AI/ML workstation specified for local model training, with quiet cooling and a clean build.

Client type
Concept build
Category
Hardware
Role
Design + engineering
Year
2026
Open the working demo
Vanta AI workstation build concept — component layout

00 / Overview

Overview

What does a machine built for one person’s actual workload look like?

Technology
  • AMD Threadripper
  • Dual NVIDIA RTX
  • 128GB ECC
  • NVMe RAID
  • Custom loop
Delivered
  • Dual-GPU platform with full-bandwidth lanes
  • ECC memory and NVMe RAID
  • Quiet sustained-load cooling
  • Linux ML environment setup

01 / The problem

We didn't start with the screen.

Training and fine-tuning models locally needs GPU memory, sustained cooling and fast storage — and most consumer builds throttle or run loud after ten minutes.

02 / The approach

We started with the system.

We specified around sustained load rather than peak benchmarks: a platform with enough PCIe lanes for two GPUs, ECC memory, NVMe RAID for datasets and a cooling design that stays quiet under hours of full utilisation.

  1. SYS / 01Component selection from the workload back: memory, GPU, storage, then everything else
  2. SYS / 02Thermals and acoustics measured under sustained load
  3. SYS / 03Clean cable work and documentation, so it can be serviced later

03 / The interface

Then we built the interface.

A workstation build specification with alternatives at three budgets, an assembly and cabling plan, a thermal validation procedure and a Linux setup optimised for ML workloads.

Use it yourself — open the demo

  • Vanta workstation build concept
    Vanta workstation build concept01 / 02
  • Vanta Workstation — the configurator on the Max tier with part swaps
    Vanta Workstation — the configurator on the Max tier with part swaps02 / 02

04 / The build

The build

What it does

  • Dual-GPU platform with full-bandwidth lanes
  • ECC memory and NVMe RAID
  • Quiet sustained-load cooling
  • Linux ML environment setup
  • Thermal and stability validation

Built with

  • AMD Threadripper
  • Dual NVIDIA RTX
  • 128GB ECC
  • NVMe RAID
  • Custom loop

05 / The result

The result

A concept build has no client numbers, and we would rather show none than invent some.

What it proves is the approach: problem first, system second, interface last. Results sections appear only on delivered client work, with verifiable outcomes.

Could we build yours?

Next project

Ember & Oak

A restaurant website with menu, reservations and a visual identity that matches the room.