AMD Ryzen AI 9
The roomier of the two: the most memory, for a small model plus a large document set or several models loaded side by side.
- Cores12
- Memory128 GB
- Graphics on the chipAMD Radeon 890M
- AI compute, CPU + GPU + NPU80 TOPS
AI inference · Choosing the hardware
Ryzen AI vs dedicated GPU: which one does your model need?
Ryzen AI vs dedicated GPU comes down to two things: how big your model is and how many people ask it questions at once. A small quantised model for one person, an internal tool or a nightly batch runs well on a Ryzen AI server from its own system memory, with no GPU sitting idle on your bill. You need a dedicated GPU when the model outgrows that memory, a crowd starts asking at the same time, you begin fine-tuning, or fast replies are what your customers pay for.
| Check | Ryzen AI is enough | Rent a GPU |
|---|---|---|
| 01Model size | Fits in system memory with room to spare | Bigger than memory, or growing |
| 02People at once | One person or a small team | Many users at the same time |
| 03The job | Serving answers, batch work | Fine-tuning or training |
| 04Reply speed | Seconds are fine | Every millisecond shows |
01
What a Ryzen AI server already does without a GPU
Before you pay for a GPU, check what you get without one. Each Ryzen AI server is a whole machine of your own: an AMD processor with Radeon graphics and an NPU on the same chip, all reading the machine's own memory. For a lot of inference work, that is the whole job.
AMD Ryzen AI Max+ 395
The faster chip: more cores and a much larger Radeon, for when replies need to come back quicker on a mid-size model.
- Cores16
- Memory64 GB
- Graphics on the chipAMD Radeon 8060S
- AI compute, CPU + GPU + NPU126 TOPS
- The whole machine is yours, with no other tenant on it
- Full root access and your own runtime, on your own versions
- NVMe storage for models, documents and checkpoints
- A 1 Gbps port that nothing meters, plus an IPv4 address
- Help on live chat: our AI replies in seconds and can check your server, and a person joins when you ask
- No slot for a GPU: if you need one later, you move to a GPU server
02
Do I need a GPU for AI inference? Four signs you do
Not always. Most small models answer fine without one. A dedicated GPU is worth its price once one of these four things happens to you.
01Model size
Your model no longer fits in memory
A model that does not fit does not run slowly. It does not load at all. Once the weights plus room for conversations pass what the machine holds, you need a GPU chosen by its memory first. Our LLM hosting guide works out how much memory your model needs.
02People at once
Lots of people ask at the same time
Answering many users together is what a GPU's fast memory and batching are built for. When questions start to queue up on a Ryzen AI server, that queue is your sign to move to a dedicated GPU server.
03The job
You start fine-tuning
Training keeps the weights, the gradients and the optimiser state in memory at the same time, several times what serving needs. That is GPU work, and for bigger models it can mean a server with several GPUs.
04Reply speed
Fast replies are part of your product
If your customers feel every pause, a GPU's faster memory is the real fix. A Ryzen AI server suits jobs where an answer in a few seconds is fine: drafting, summaries, tagging, internal tools.
03
CPU inference vs a GPU server: memory decides what loads, speed decides who waits
On a Ryzen AI server, space is rarely the problem. Both machines hold the weights of every model size below at 4-bit, with memory left for the system. The limit is speed: every word the model writes reads the weights again, and system memory reads more slowly than the memory on a GPU. One person on a small model will not notice. A queue of people will.
| Model size | Weights at 4-bit | Weights at 8-bit | Weights at 16-bit |
|---|---|---|---|
| 7-8B | 5 GB | 8 GB | 16 GB |
| 13-14B | 8 GB | 14 GB | 28 GB |
| 30-34B | 19 GB | 34 GB | 68 GB |
| 70B | 38 GB | 70 GB | 140 GB |
These are the weights only. Budget a quarter to a half again on top for conversations and the runtime. For comparison, the Ryzen AI servers carry 128 GB and 64 GB of memory. We publish no speed figure for these machines, because it changes with every model and runtime. Test your own model on one, or ask us on chat what to expect.
04
NPU vs GPU inference, and the options in between
The NPU on a Ryzen AI chip is real, but it is not the part most people end up using. Here is what each option is good for.
ANPU
The Ryzen AI NPU
It runs models through AMD's own Ryzen AI software. Most open-model runtimes people deploy use the processor and the built-in Radeon instead. Rent one by the month to test the NPU tools on real hardware before you commit to more.
BUnified memory
The DGX Spark
The same idea as Ryzen AI at the top end: a GPU that can address 128 GB of memory in one compact machine. Worth a look when you want to fine-tune or run a large model on an NVIDIA GPU but still serve a small audience. See the DGX Spark server.
CDedicated GPU
A GPU server
A dedicated GPU has its own fast memory, so it answers more people faster and is the only real option for fine-tuning. Every GPU we fit is priced against the exact server it goes in on our GPU servers page.
05
Questions people ask before choosing
01Do I need a GPU for AI inference?
Not for a small quantised model with light traffic. A Ryzen AI server runs it from system memory at a pace that suits one person, an internal tool or a batch queue. You need a dedicated GPU when the model is bigger than the memory, many people use it at once, you fine-tune, or reply speed matters to your customers.
02How does a Ryzen AI server compare to a GPU server?
They do the same job with different memory. A Ryzen AI server keeps the model in system memory, which is plentiful for the price but slower to read. A GPU server keeps it in the GPU's own memory, which is scarcer and much faster. Your model size and your traffic decide between them.
03Will the NPU run my model?
It can, through AMD's Ryzen AI software. Most open-model runtimes use the processor and the built-in Radeon graphics instead, so plan on those unless your tools target the NPU. It is on the chip either way, and nothing stops you using it.
04Can I add a GPU to a Ryzen AI server later?
No. These machines have no slot for a GPU. If you outgrow one, you move to a GPU server instead, and we can help you plan that move on chat.
05Can you help me move my setup from one to the other?
Yes. Bring your containers, models and data, and tell us on live chat what you run before you order. Our AI assistant replies in seconds and can look at your account and servers, and a person from our team takes over whenever you ask.
06Where are the prices?
The Ryzen AI servers are priced live on the Ryzen AI page, and every GPU is priced with the server it goes in on the GPU servers page. Keeping them there means you always see today's price.
Still not sure which side you are on?
Tell us your model and how many people will use it. Our live chat answers in seconds and will point you to a Ryzen AI server or a GPU server, and a person is there if you want one.