What unified memory is
In a classic computer, the processor has its memory and the graphics card has its own, and a model has to fit in the card’s. In a processor with unified memory, both share the same 64 or 128 GB with high bandwidth: the integrated GPU can use almost all of that memory for the model. The result is a 10-litre unit, quiet and low-power, that runs models that do not fit in a 16 or 24 GB graphics card.
What it is for
For private assistants that summarise, draft and search the company’s documents; for language models of 30 to 70 billion parameters, quantised, that answer at reading speed; for offline vision and speech models; and for development and testing before moving to a server. All without data leaving the company and with the power draw of a large laptop.
What it is not for
For training models or serving many people at once: there compute power and the bandwidth of a dedicated GPU rule, and unified memory falls several times behind. Nor for generating images or video at production pace. In those cases, a workstation with an RTX PRO or an AI server.
How much memory you need
| Model | Memory it takes | Machine |
|---|---|---|
| 7 to 8 billion parameters | 6 to 8 GB | Any machine with 16 GB or an 8 GB graphics card |
| 14 to 32 billion | 10 to 24 GB | 32 GB unified or a 24 GB graphics card |
| 70 billion | 40 to 48 GB | 64 GB unified |
| Several models at once or a very long context | 64 to 100 GB | 128 GB unified |
Leave headroom: the system, the conversation context and the documents loaded also take memory.
The form factor: 10 litres, quiet and wherever it is needed
A compact unit with unified memory fits on a desk, in an on-board cabinet or in a reception counter, with 12 to 24 V power if needed and, in the industrial version, fanless. We deliver it with the system and the inference engine installed, tested for 24 hours, and with the models you choose loaded. The price is quoted according to memory and form factor.
Related hardware and guides
Frequently asked questions
Is it slower than a dedicated GPU?
In single-user response speed with a model that fits in both, the dedicated GPU is faster. The advantage of unified memory is that it runs models that do not fit in a 16 or 24 GB card, with far less power draw and noise.
How many people can use it at once?
A few. It is designed for a small team or a single desk, not to serve the whole company. For dozens of users, an AI server with a dedicated GPU.
Which models can I run?
Open language, vision and speech models, quantised: with 64 GB, models of up to 70 billion parameters; with 128 GB, several at once or very long contexts. We leave them loaded and tested.
Can it go on a boat or in a vehicle?
Yes. The industrial version is fanless, runs on 12 to 24 V, is vibration-tested and fits in shallow cabinets.
Where will your computer work?
Tell us the location, the power available and what it has to do. We will propose the compact unit, with tests and a price, within 12 working hours.