Think AI UltraNode: A Rack of Compute in One Node
Seven accelerators held at peak clocks continuously, 672 GB of unified memory, and a self-contained 6000 W liquid-cooling system — with no facility water.
Some organizations are not satisfied with running models; they want to teach them. Think AI UltraNode — available from Linamara — is designed for exactly that: continuous training on the organization's documents and terminology while serving users at the same time, on a single node that stands on an office or lab floor.
Contact Us Now
Contact us now if you have any inquiries or wish to learn more about our integrated services and solutions.
What is the Think AI UltraNode?
Key figures
Who is it for?
- Large institutions: archives of millions of documents and large numbers of users.
- Continuous training: regularly fine-tuning models on new documents and institutional terminology.
- Large models: open models of a size that fits within 672 GB of memory.
- Innovation and research centers: experimenting with, evaluating and running models in one environment.
Technical specifications
Train while serving
In conventional architectures, training needs an environment separate from serving, so part of the hardware sits idle in each. On the UltraNode, the ILM platform places serving jobs and fine-tuning jobs (SFT, LoRA and QLoRA) side by side across the seven accelerators, according to each accelerator’s available memory and thermal headroom — so the model learns your organization’s terminology by night and serves your staff by day, on the same node.
- Seven accelerators at peak clocks, continuously, thanks to self-contained 6000 W liquid cooling.
- 1 TB of system memory to keep the accelerators fed with data without bottlenecks.
- 3% sustained utilization under multi-model load — per the manufacturer’s measurements.
The Linamara systems it runs
- AI assistant: a model fine-tuned on your organization’s language and internal terminology.
- Talk to your data: querying very large archives with many concurrent users.
- AI document processing: reading and classification models that keep improving from user corrections.
Supply, installation and support
Linamara handles sizing, supply, installation, connection to its AI systems and to Arshef, training for your team and direct local support — backed by the manufacturer’s warranty: 3 years on hardware and 1 year of software support, with Premium and Mission-Critical service levels on request.
Teach AI your organization’s language
We will prepare a proposal for an UltraNode together with a plan for fine-tuning models on your documents and a clear estimate of the workloads.
▶ Request a quote | Book a sizing session
Do you have any inquiries?
Contact us now if you have any questions or wish to learn more about our integrated services and solutions. Our professional team is ready to provide support and answer all your inquiries with speed and efficiency, ensuring you receive the best possible experience and service.
Frequently asked questions
Q1. What is the difference between serving and fine-tuning?
Serving (inference) is the model answering users’ questions; fine-tuning is training it on examples from your documents so it masters their terminology and style. The UltraNode does both on the same node.
Q2. Does the training data stay inside the organization?
Yes. Training runs on the node inside your network, so neither the documents nor the resulting weights leave the organization — and the fine-tuned model is yours.
Q3. Where is the UltraNode installed?
It comes as a tower that stands on an office or lab floor, or as a rackmount unit — and neither needs facility cooling water.
Q4. When should we choose the UltraNode over the SuperNode?
When continuous model training is part of the plan, or when your models and user numbers exceed what 384 GB of memory can hold.