GPT-OSS 20B and 120B Models on AMD Ryzen AI Processors (www.amd.com)

submitted 2 days ago by sith@lemmy.zip to c/localllama@sh.itjust.works

11 comments fedilink hide all child comments

you are viewing a single comment's thread
view the rest of the comments

[-] tal@lemmy.today 2 points 2 days ago* (last edited 2 days ago)

It looks like they're running LM Studio on an AI Max 395+. I'm confident that you can do the same on Linux. Here's a post on Framework's blog talking about comparable hardware, the Framework Desktop (also 128GB Ryzen AI Max 395+).

https://frame.work/blog/using-a-framework-desktop-for-local-ai

The simplest application we’ve found to get started with for the text and code generation AI use case is LM Studio. It’s built on top of llama.cpp, which is an extremely powerful and extensible open source inference library that we’ll go into in a future guide. LM Studio packages it up in a user-friendly application that runs on both Windows and Linux. With this guide, we’ll focus on Windows, but the same settings will work on Linux. Note that as of June 2025, inference runs about 20% faster on Fedora 42 than on Windows 11.

EDIT: Also, getting a bit off the point, I'd guess that something like the Framework Desktop (not necessarily specifically a Framework Desktop, but at least something in a more desktop/server-like mini-ITX form factor) is probably preferable to a laptop-like ROG Flow Z13 if the goal is a pure AI box, because it'll be able to dissipate heat better. It looks like Asus capped the Z13's power usage, I imagine because of the heat (and maybe battery, though I can't imagine running LLM stuff on a Z13 on battery would be very practical):

https://old.reddit.com/r/FlowZ13/comments/1ki68oj/a_bit_disappointed_the_new_z13_2025_performance/

Z13 2025 TDP is fixed to 35W-55W quiet mode, 45W~70W performance mode, 60W-86W turbo mode and up-to 93W manual mode

https://frame.work/blog/introducing-the-framework-desktop

In a desktop form factor, we get to unlock every bit of its performance with 120W sustained power and 140W boost while staying quiet and cool.

EDIT2: Also, I don't think that Windows is a great fit with the hardware, if the aim is a dedicated AI box. Like, the reason you'd want to get an AI Max for generative AI purposes in 2025 is because you want something that isn't insanely expensive that can have a whole lot of memory hooked up to a GPU. People noticed that "unified" systems that have shared GPU and CPU memory is a good way to achieve that. Okay...but if you do that, then what you're paying for is a lot of GPU-accessible memory. But then any other software that you're running on the machine is consuming some of that memory. You probably want to be able to run the thing as stripped-down as possible, and it's gonna be rather more practical to do that with a minimal Linux box than with a Windows 11 machine.

[-] infinitevalence@discuss.online 4 points 2 days ago* (last edited 1 day ago)

Just ran it on my Framework 13 w/ Ryzen AI 9 HX 370. Its not able to use the NPU at all which is as expected in Linux. :(

Still getting 20 tokens/sec is not bad off the GPU/CPU and 2x what I was seeing on Gemma 3 12b.

I have a Radeon RX 9060 XT 16GB which I may try to plug in and bench too.

This is a nice uplift in performance but its still not utilizing dedicated hardware.

Ryzen AI 9 HX 370 w/ 870M + 64GB gpt-oss-20b:

 "stats": {
    "stopReason": "userStopped",
    "tokensPerSecond": 19.52169768309363,
    "numGpuLayers": -1,
    "timeToFirstTokenSec": 7.363,
    "totalTimeSec": 24.127,
    "promptTokensCount": 1389,
    "predictedTokensCount": 471,
    "totalTokensCount": 1860

Ryzen 9 3900x w/ RX 9060 XT 16GB + 128GB gpt-oss-20b

"stats": {
    "stopReason": "eosFound",
    "tokensPerSecond": 21.621644364467713,
    "numGpuLayers": -1,
    "timeToFirstTokenSec": 0.365,
    "totalTimeSec": 15.124,
    "promptTokensCount": 118,
    "predictedTokensCount": 327,
    "totalTokensCount": 445

Ryzen AI 9 HX 370 w/ 870M + 64GB gemma-3-12b

 "stats": {
    "stopReason": "eosFound",
    "tokensPerSecond": 7.866126447290582,
    "numGpuLayers": -1,
    "timeToFirstTokenSec": 12.937,
    "totalTimeSec": 12.458,
    "promptTokensCount": 1366,
    "predictedTokensCount": 98,
    "totalTokensCount": 1464

Ryzen 9 3900x w/ RX 9060 XT 16GB + 128GB gemma-3-12b

"stats": {
    "stopReason": "eosFound",
    "tokensPerSecond": 27.62905167372976,
    "numGpuLayers": -1,
    "timeToFirstTokenSec": 0.31,
    "totalTimeSec": 19.219,
    "promptTokensCount": 2975,
    "predictedTokensCount": 531,
    "totalTokensCount": 3506

[-] hendrik@palaver.p3x.de 1 points 2 days ago* (last edited 2 days ago)

Thanks for the numbers. Btw, I think a NPU can't run large language models in the first place. They're meant for things like blur the background in video conferences, or help with speech recognition or such very specific smaller tasks. They only have some tens or hundreds of megabytes of memory, so a LLM/chatbot won't fit. The main thing that makes LLM inference faster is memory (RAM) bandwith and speed.

[-] infinitevalence@discuss.online 2 points 1 day ago

I added a few more numbers. I may pull my old MI25 out of mothballs and bench that to.

this post was submitted on 14 Aug 2025

15 points (100.0% liked)

LocalLLaMA

3540 readers

12 users here now

Welcome to LocalLLaMA! Here we discuss running and developing machine learning models at home. Lets explore cutting edge open source neural network technology together.

Get support from the community! Ask questions, share prompts, discuss benchmarks, get hyped at the latest and greatest model releases! Enjoy talking about our awesome hobby.

As ambassadors of the self-hosting machine learning community, we strive to support each other and share our enthusiasm in a positive constructive way.

Rules:

Rule 1 - No harassment or personal character attacks of community members. I.E no namecalling, no generalizing entire groups of people that make up our community, no baseless personal insults.

Rule 2 - No comparing artificial intelligence/machine learning models to cryptocurrency. I.E no comparing the usefulness of models to that of NFTs, no comparing the resource usage required to train a model is anything close to maintaining a blockchain/ mining for crypto, no implying its just a fad/bubble that will leave people with nothing of value when it burst.

Rule 3 - No comparing artificial intelligence/machine learning to simple text prediction algorithms. I.E statements such as "llms are basically just simple text predictions like what your phone keyboard autocorrect uses, and they're still using the same algorithms since <over 10 years ago>.

Rule 4 - No implying that models are devoid of purpose or potential for enriching peoples lives.

founded 2 years ago

MODERATORS

SkySyrup@sh.itjust.works

pax@sh.itjust.works

noneabove1182@sh.itjust.works

Smokeydope@lemmy.world

MonsterBug@sh.itjust.works