GMKtec EVO-X1 MFF AI Mini PC: Ryzen AI 9 HX 370, 32GB LPDDR5x, 2TB SSD, OCulink $699.99

jasoncwlee

Sep 20, 2025 4:41 AM

438 Comments

Joined Feb 2008

jasoncwleeSep 20, 2025 4:41 AM

438 Comments

Is this good for running AI model locally using unified memory?

1

Dr.W

Original Poster

Pro

Sep 20, 2025 5:08 AM

17.8K Comments

Joined Nov 2020

Dr.W

Original Poster

Pro

Sep 20, 2025 5:08 AM

17.8K Comments

Quote from jasoncwlee

:

Is this good for running AI model locally using unified memory?

Yes, it's a very good choice to locally run AI models if you optimize (use quantization, efficient frameworks, keep models reasonable in size). It's especially good for inference, testing, and lighter workloads. For large models or many simultaneous sessions, a discrete GPU rig might do more, but this is solid for what many users want locally.

3

1

insaneXproval

Sep 20, 2025 1:08 PM

835 Comments

Joined Apr 2014

insaneXprovalSep 20, 2025 1:08 PM

835 Comments

Quote from Dr.W

:

Yes, it's a very good choice to locally run AI models if you optimize (use quantization, efficient frameworks, keep models reasonable in size). It's especially good for inference, testing, and lighter workloads. For large models or many simultaneous sessions, a discrete GPU rig might do more, but this is solid for what many users want locally.

That Oculink port can help with the dedicated GPU part. And Oculink generally has better performance for eGPUs than USB4.

1

mesowetodit

Sep 20, 2025 4:32 PM

694 Comments

Joined Jul 2007

mesowetoditSep 20, 2025 4:32 PM

694 Comments

man, i feel so old...there USB4 now??? i can't keep up with this...wheres my ps1...

2

1

2

cuoreesitante

Sep 20, 2025 5:40 PM

2.1K Comments

Joined Aug 2008

cuoreesitanteSep 20, 2025 5:40 PM

2.1K Comments

Quote from Dr.W

:

Yes, it's a very good choice to locally run AI models if you optimize (use quantization, efficient frameworks, keep models reasonable in size). It's especially good for inference, testing, and lighter workloads. For large models or many simultaneous sessions, a discrete GPU rig might do more, but this is solid for what many users want locally.

You'd be pretty limited as to what models you can run thou. Considering the ram is not upgradable I'd opt for at least 64gb if you want to run local models with the next few years in mind.

1

sr27

Sep 20, 2025 11:25 PM

1.3K Comments

Joined Apr 2008

sr27Sep 20, 2025 11:25 PM

1.3K Comments

Quote from cuoreesitante

:

Quote from Dr.W

:
Yes, it's a very good choice to locally run AI models if you optimize (use quantization, efficient frameworks, keep models reasonable in size). It's especially good for inference, testing, and lighter workloads. For large models or many simultaneous sessions, a discrete GPU rig might do more, but this is solid for what many users want locally.

You'd be pretty limited as to what models you can run thou. Considering the ram is not upgradable I'd opt for at least 64gb if you want to run local models with the next few years in mind.

I am no expert, but this seems like a bad config to run AI. The soldered ram is a big blocker. Plus you can only connect one gpu using oculink. May be squeeze in another on usb4. Either way this can probably run small models and even be used for tuning small models. But not suitable above hobby level AI activities.

1

2

RyanL

Sep 20, 2025 11:31 PM

6.6K Comments

Joined Feb 2008

RyanLSep 20, 2025 11:31 PM

6.6K Comments

I grabbed one of these a couple of days back for my son's birthday. Seemed like a the best deal I could find on mini pc with 890m GPU. It was the last one in parkville MD though (not sure if other stores have more stock).

1

lolrofllol

Sep 20, 2025 11:46 PM

559 Comments

Joined Sep 2011

lolrofllolSep 20, 2025 11:46 PM

559 Comments

Quote from sr27

:

I am no expert, but this seems like a bad config to run AI. The soldered ram is a big blocker. Plus you can only connect one gpu using oculink. May be squeeze in another on usb4. Either way this can probably run small models and even be used for tuning small models. But not suitable above hobby level AI activities.

Soldered memory is preferable because LPDDR5X has more bandwidth than socketed DDR5. Although, LPDDR5x-7500 is not the fastest spec available. These unified memory designs benefit heavily from the additional memory bandwidth of soldered RAM, and AI inference tends to be bound by bandwidth more than compute in these situations.

Besides, I'm pretty sure that Ryzen AI 300 series come ONLY in soldered memory designs - unsure if that's a technical requirement or just arbitrarily set by AMD, but it does make sense. Even in gaming scenario, we'd probably see LPDDR5X-7500 perform 20-30% better than socketed DDR5-5600.

1

lennonst

Sep 21, 2025 10:21 AM

2.9K Comments

Joined Jun 2012

lennonstSep 21, 2025 10:21 AM

2.9K Comments

Quote from Dr.W

:

Yes, it's a very good choice to locally run AI models if you optimize (use quantization, efficient frameworks, keep models reasonable in size). It's especially good for inference, testing, and lighter workloads. For large models or many simultaneous sessions, a discrete GPU rig might do more, but this is solid for what many users want locally.

You could just say youre not sure, since clearly that is the case. 32gb is garbage for running useful models.

2

lennonst

Sep 21, 2025 10:24 AM

2.9K Comments

Joined Jun 2012

lennonstSep 21, 2025 10:24 AM

2.9K Comments

Quote from lolrofllol

:

Soldered memory is preferable because LPDDR5X has more bandwidth than socketed DDR5. Although, LPDDR5x-7500 is not the fastest spec available. These unified memory designs benefit heavily from the additional memory bandwidth of soldered RAM, and AI inference tends to be bound by bandwidth more than compute in these situations.

Besides, I'm pretty sure that Ryzen AI 300 series come ONLY in soldered memory designs - unsure if that's a technical requirement or just arbitrarily set by AMD, but it does make sense. Even in gaming scenario, we'd probably see LPDDR5X-7500 perform 20-30% better than socketed DDR5-5600.

You have no idea what he was posting about, lol. A measly 32gb of soldered ram is garbage for this thing as it will be the limiting factor in running large LLM models which is the entire point of his post.

2

lolrofllol

Sep 22, 2025 4:56 PM

559 Comments

Joined Sep 2011

lolrofllolSep 22, 2025 4:56 PM

559 Comments

Quote from lennonst

:

You have no idea what he was posting about, lol. A measly 32gb of soldered ram is garbage for this thing as it will be the limiting factor in running large LLM models which is the entire point of his post.

You clearly have never tried to run any models on your own hardware. More capacity is worthless without sufficient memory bandwidth, and on Strix Point, we're severely bandwidth limited. Soldered LPDDR5x in this case will provide somewhere around 20-30% more bandwidth than socketed DDR5, hence the preference.

Sure, we can technically run larger models with more socketed RAM, but it'll be so painfully slow that it's not actually useful in any sort of real-time use case.

32GB is plenty enough to run the sort of models that can actually run fast enough on this level of compute and bandwidth.

lennonst

Sep 22, 2025 5:38 PM

2.9K Comments

Joined Jun 2012

lennonstSep 22, 2025 5:38 PM

2.9K Comments

Quote from lolrofllol

:

You clearly have never tried to run any models on your own hardware. More capacity is worthless without sufficient memory bandwidth, and on Strix Point, we're severely bandwidth limited. Soldered LPDDR5x in this case will provide somewhere around 20-30% more bandwidth than socketed DDR5, hence the preference.

Sure, we can technically run larger models with more socketed RAM, but it'll be so painfully slow that it's not actually useful in any sort of real-time use case.

32GB is plenty enough to run the sort of models that can actually run fast enough on this level of compute and bandwidth.

His point has still gone over your head, no one is arguing soldered ram isn't better.

But you clearly have never run a useful model locally, as 32gb is only enough to run gimmick level LLMs.

DJ-GMONEY

Sep 24, 2025 5:24 PM

110 Comments

Joined Nov 2009

DJ-GMONEYSep 24, 2025 5:24 PM

110 Comments

A friend of mine who works in Cyber Security told me to stay away from these Mini PCs. They can have embedded Spyware. Not sure how true that is.

IanB8513

Sep 24, 2025 10:50 PM

180 Comments

Joined Apr 2015

IanB8513Sep 24, 2025 10:50 PM

180 Comments

Quote from lennonst

:

His point has still gone over your head, no one is arguing soldered ram isn't better. But you clearly have never run a useful model locally, as 32gb is only enough to run gimmick level LLMs.

Even the 70-90b models are gimmicky compared to frontier models and run like trash locally compared to open router or groq and with a fraction of the context.So if you're saying there's actually a good reason to have 64gb on this machine there's most likely someone who could say the same about 32.If you don't know how to effectively leverage 32gb that says more about your experience in this area compared to someone who says they can.

GMKtec EVO-X1 MFF AI Mini PC: Ryzen AI 9 HX 370, 32GB LPDDR5x, 2TB SSD, OCulink $699.99

$700

$1,600

Community Voting

Join The Conversation

14 Comments

Join The Conversation

Related Searches