tandemcompute standard node

NVIDIA RTX PRO 6000 Blackwell.

NVIDIA’s all-round Blackwell data-centre GPU, run liquid-cooled on tandemcompute infrastructure in Ireland.

platform overview

GPU compute in a flexible form factor.

The RTX PRO 6000 delivers high-performance GPU compute in a standard PCIe form factor, a strong fit for AI inference, fine-tuning, rendering and visualisation. It gives teams serious GPU performance without the infrastructure complexity of HGX-class platforms, and puts twice the GPUs into each node as a flagship chip would, for less capital.

platform specs

NVIDIA RTX PRO 6000 Server Edition GPU.

A PCIe-based GPU built on NVIDIA’s Blackwell architecture. Strong compute and graphics performance in a standard form factor, suitable for a wide range of workloads.

compute

Blackwell architecture

24,064 CUDA cores and fifth-generation Tensor Cores with FP4, around 4 PFLOPS of AI compute per GPU. Fourth-generation RT Cores for rendering.

memory

96 GB GDDR7 memory

ECC memory at 1.6 TB/s. Most production models fit on one card without sharding.

form factor

PCIe, liquid-cooled

Dual-slot PCIe Gen 5, 600 W per GPU, direct-to-chip liquid cooling in our nodes. Up to four MIG instances per card.

how it is sold

Single GPU to whole nodes

From a single GPU to whole nodes. Reserved by the month or on demand by the hour.

Specifications per NVIDIA’s RTX PRO 6000 Blackwell Server Edition datasheet.

use cases

Common workloads.

inference

Inference and agents

Serve 30B to 235B-class open models on one to four cards. 96 GB holds the weights and a long context; FP4 keeps tokens cheap.

rendering

Rendering and visualisation

ComfyUI and diffusion pipelines, video generation, Blender and Unreal render queues. Hardware ray tracing and the memory for large scenes.

training

Fine-tuning and vision

LoRA and full fine-tunes of mid-size models, computer-vision training and batch embedding runs that would otherwise wait for scarce flagship GPUs.

why tandem

Hosted in Ireland

Physically in Ireland, EU jurisdiction, tandem-controlled. Data locality is a fact of the site, not a contract clause.

Single GPU to whole nodes

Sold singly, in twos, fours or eights, or as whole nodes. No minimum term on demand.

Heat returned

Liquid-cooled in a closed loop; the heat is metered and returned to the host building.

Someone who answers

A named engineer by email and phone. Small enough to know your workload.