The server · local AI
Prototype · under reviewA server with 1 to 8 NVIDIA RTX PRO 6000 GPUs and liquid cooling, installed in your office. 3XS Systems assembles and tests it in the United Kingdom; we size it for your company and leave it up and running.
01 · GPU
The GPUs are what run the model. Each one has 96 GB of GDDR7 memory with error correction and 24,064 CUDA cores. We fit as many as your company needs, depending on how many people use it and how large the model is.
With 8 GPUs, the platform switches to an AMD EPYC processor.
02 · Liquid cooling
Each GPU can draw up to 600 W. A water loop carries that heat to the radiators, so the server is far quieter than an air-cooled one and can live in an office.
03 · Processor
It coordinates the work and feeds data to the GPUs. Its platform has 128 PCIe 5.0 lanes, enough for every GPU to run at full speed.
Reference: Threadripper PRO 9975WX. Available with up to 96 cores.
04 · Memory
Eight 64 GB modules at 5,600 MHz with error correction, the memory used in servers that cannot fail. It sustains the system and prepares the data headed for the GPUs.
Configurable from 128 GB.
05 · Storage
A Samsung 9100 PRO SSD holds the system, the models and your company's memory. It reads at 14,700 MB/s: a 60 GB model loads in seconds.
06 · Power
Two 3,000 W power supplies under the bottom cover feed the GPUs and the rest of the machine. With 7 GPUs at full load, peak draw is around 5 kW, which is why it connects to two independent electrical circuits.
We check this during the site visit, before installing.
07 · Assembled and tested
3XS Systems builds it by hand in the United Kingdom and puts it through a load test with AI workloads before shipping. It comes with Ubuntu 24.04, NVIDIA CUDA drivers and Docker. We install the model and Kelvo Desktop, and leave it running on your network.
01 Configure
The number of GPUs sets how much memory the model has and how much work it can handle at once. Pick a number to see what changes.
NVIDIA RTX PRO 6000 GPUs · 96 GB each
Several areas at once, or a large model with a long context for lengthy documents.
Reference estimates. The model size assumes 8-bit weights and leaves about 25% of the memory for context. We set the final configuration in the assessment, measuring with your real tasks.
02 Supplier
It isn't a prototype: it's a machine sold today, assembled and tested, by an integrator with more than 20 years building AI workstations and servers. We choose the configuration, import it and leave it running with the model and Kelvo Desktop.
Their liquid-cooled AI workstation for up to 7 GPUs, designed to work in an office rather than a data center.
Professional-grade parts: error-correcting memory, a workstation motherboard and a PCIe 5.0 SSD.
Each machine is built by hand and tested with deep learning workloads at the factory, with the operating system and drivers already installed.
We measure your tasks and choose the model, the number of GPUs and the memory.
We finalize the configuration and send you the quote.
5 to 10 business days of assembly and load testing in the United Kingdom.
We bring it to Chile and coordinate delivery.
We leave it on your network with the model and Kelvo Desktop running.
03 Spec sheet
It's the one in the drawing: seven GPUs, the most this platform takes. The right-hand column shows what can be adjusted.
| Component | Reference | Options |
|---|---|---|
| GPU | 7 × NVIDIA RTX PRO 6000 · 96 GB GDDR7 ECC each (672 GB in total) | From 1 to 8* |
| Processor | AMD Ryzen™ Threadripper™ PRO 9975WX · 32 cores / 64 threads · 4.0 to 5.4 GHz | Up to 96 cores |
| Motherboard | ASUS Pro WS WRX90E-SAGE SE · PCIe 5.0 | |
| Memory | 512 GB DDR5 ECC at 5,600 MHz (8 × 64 GB) | From 128 GB to 1 TB |
| Storage | 4 TB Samsung 9100 PRO · NVMe PCIe 5.0 · 14,700 MB/s read | 4 M.2 slots plus bays for SSDs and HDDs |
| Cooling | Custom liquid cooling · 2 independent loops with dual pumps · quick-disconnect fittings | |
| Power | 2 × 3,000 W power supplies (6,000 W) · peak draw close to 5 kW | |
| Networking | 10 GbE + 1 GbE + remote management port (IPMI) | Faster NVIDIA network cards |
| Chassis | Corsair 9000D · 698 × 698 × 307 mm (height × depth × width) | |
| Base software | Ubuntu 24.04 · NVIDIA CUDA · Docker and NVIDIA Container Toolkit | Plus the model and Kelvo Desktop, which we install |
| Warranty | 3 years on parts and labor |
* The 8-GPU configuration uses an AMD EPYC processor.
04 In your office
A large tower that sits on the floor, with clearance around it so the radiators can breathe.
With 7 GPUs at full load it draws close to 5 kW. Configurations with fewer GPUs draw considerably less.
All the power it draws ends up as heat in the room, so a ventilated or air-conditioned space is best.
It connects over 10 GbE or 1 GbE. Internet access for installation, updates and web search. The model and your documents stay on your server.
We check these four points in a site visit before installing.
We measure with your real tasks, choose the model and send you a quote for the server with installation included.