An eight-card Instinct server
Eight MI355X cards give about 2.3 TB of memory in one server, for trillion-parameter models and dozens of concurrent users.
AMD AI servers for large models
AMD Instinct cards are among the highest-memory cards on the market; the MI355X has 288 GB per card, so a large model fits on fewer cards. Alongside them we supply and set up Radeon cards for workstations and small Ryzen AI Max computers.
Why AMD
Models run on AMD through ROCm. Its version 10.1 (October 2026) officially supports vLLM, SGLang and PyTorch, and Ollama runs on Radeon and Instinct cards too. AMD is a good choice for large models that need a lot of memory, but some new models and optimisations reach it later than NVIDIA, so before you buy we test your model on the same hardware.
Which tier for which job
A PCIe card in an ordinary server
The MI350P, with 144 GB of HBM3E, fits ordinary air-cooled servers and needs no special chassis.
A workstation with Radeon cards
Radeon AI PRO R9700 with 32 GB or Radeon PRO W7900 with 48 GB, for mid-size models in one department.
A small Ryzen AI Max computer
Ryzen AI Max+ 395 with 128 GB of shared memory in a book-sized computer; AMD says it runs models of up to 200 billion parameters. The newer Ryzen AI Max+ PRO 495 goes up to 192 GB.
Instinct cards and servers
Specifications from AMD's official pages, reviewed on 22 October 2026.
| Product | Memory per card | Bandwidth | In an eight-card server | What fits |
|---|---|---|---|---|
| Instinct MI300X | 192 GB HBM3 | 5.3 TB/s | 1.5 TB | A model of about 150B at 8-bit per card |
| Instinct MI325X | 256 GB HBM3E | 6 TB/s | 2 TB | A model of about 200B at 8-bit per card |
| Instinct MI350X and MI355X | 288 GB HBM3E | 8 TB/s | 2.3 TB | Trillion-parameter models at 8-bit in one server |
| Instinct MI350P (PCIe card) | 144 GB HBM3E | 4 TB/s | Depends on the server | A model of about 120B at 8-bit per card |
| Instinct MI455X and the Helios rack | 432 GB HBM4 | 23.3 TB/s | A 72-card rack with 31 TB of memory | Multi-trillion-parameter models; liquid-cooled |
The "what fits" column is our estimate, not a vendor claim: each billion parameters takes about 1 GB of memory at 8-bit and about half that at 4-bit, and room must be left for context and concurrent users. The MI350P and MI455X are newly announced; their delivery time is stated in the quote.
Workstations and small computers
Specifications from AMD's official pages, reviewed on 22 October 2026.
| Product | Memory | Bandwidth | Power | What fits |
|---|---|---|---|---|
| Radeon AI PRO R9700 | 32 GB GDDR6 | 640 GB/s | 300 W | A model up to about 30B at 4-bit; four cards give 128 GB |
| Radeon PRO W7900 and W7800 48GB | 48 GB GDDR6 | 864 GB/s | 260 to 295 W | A model of about 30B at 8-bit |
| Ryzen AI Max+ 395 | Up to 128 GB shared | About 256 GB/s (our calculation) | 45 to 120 W | Up to 200B parameters according to AMD; suits MoE models |
| Ryzen AI Max+ PRO 495 | Up to 192 GB (up to 160 GB for the GPU) | About 273 GB/s (our calculation) | 45 to 120 W | Over 300B parameters at 4-bit according to AMD |
The "what fits" column is our estimate, not a vendor claim: each billion parameters takes about 1 GB of memory at 8-bit and about half that at 4-bit, and room must be left for context and concurrent users.
The software we set up on it
| Layer | What we install and tune |
|---|---|
| Serving for many users | vLLM or SGLang, both of which batch concurrent requests together |
| Light and test serving | llama.cpp or Ollama for a single user, testing and small models |
| Programming interface | An OpenAI-compatible API, so applications that call a foreign service today connect to the local model without a rewrite |
| Assistant and documents | A Persian assistant, search over internal documents with citations (RAG), and n8n workflows |
| Sign-in and monitoring | Sign-in with company accounts, logging of questions and answers, and monitoring of GPU use, request queues and response time |
From the first call to a working system
Hardware is chosen after a model has been measured on your own work, not from a table. A quote is issued only once it is clear which model has to serve how many people, and how fast.
- 01
First conversation
We go through the task, the number of users, the data, where the hardware goes and what you already own.
- 02
Measured on your work
We test several open-weight models on samples of your real work and size memory and speed from those results.
- 03
Written quote
You get the exact configuration, the proposed model, the delivery time and the support terms in writing. Prices appear only in the quote.
- 04
Installation
We install the hardware, bring up the model and connect the assistant to your systems, then test it once with international internet cut off.
- 05
Handover and support
You get documentation and training for your team; afterwards we maintain the system and update the models under contract.
How to get in touch
Fill in the form or message us on WhatsApp. The clearer these are, the more exact the quote:
- What the model has to do (answer from documents, code, summarise, process images or speech)
- How many people use it at the same time, and what response time is acceptable
- The model or hardware family you have in mind, if you already know
- Where it goes: your server room, a data centre, or hardware you already own
- How sensitive the data is, and whether it must run with no internet at all
We reply within one business day.
Zharf
Try the ZharfAI assistant right now
Ask about Persian assistants, document automation, or how to start a pilot.
Frequently asked questions about AMD AI servers
What is ROCm, and do all models run on AMD?
ROCm is AMD's software for running AI computation on its cards, much like CUDA on NVIDIA. vLLM, SGLang and PyTorch officially run on it, and most common open models work. Some new models or specific optimisations ship for NVIDIA first, which is why we test your model on the same hardware before you buy.
MI355X or B200?
With 288 GB per card, the MI355X has more memory than the B200 (180 to 192 GB) and fits a larger model in one server. NVIDIA's B200 and B300 have wider software support. The choice depends on the model, the number of users and the software you use, and we settle it by testing on your real work.
What is Ryzen AI Max good for?
An assistant for one department, development and testing, in a small, low-power computer. Its memory bandwidth is about 256 GB/s, so it suits MoE models that read only part of their parameters for each word (such as gpt-oss-120b or Qwen3.5-122B-A10B) better than large dense models.
Do you rent GPU servers by the hour or month?
No. ZharfAI does not rent GPUs by the hour or month. There are three ways to work with us: we supply the hardware and install it on your premises, we run a dedicated machine for you and maintain it ourselves, or we set the model up on hardware you already own. If you only need a few hours of compute for a test, say so in the first conversation and we will suggest a suitable route.
How much does it cost?
We do not publish prices, because they depend on the configuration, quantity, the day's market and how it is delivered. After the first conversation and a test on your work you receive a written quote that also names the exact hardware model, the delivery time and the support terms.
Does it work without international internet?
Yes, if the model, the serving software, the assistant, user sign-in and monitoring all sit inside your network. Anything that depends on a foreign service stops during a cut, so before handover we test the system with international internet deliberately cut off.
Test the model on AMD before you buy
Tell us which model you want to run and for how many users. We measure it on AMD, and if NVIDIA or a Mac suits your work better, we will say so plainly.