You are right. Not sure if I ever want to upgrade. I have a very limited space and therefore I searched for an iGPU. At the end there are problems with minipc too when you check comments at Amazon. The user makes the end control.
So the only alternative I see, is an ITX-Mini_PC, but here you have to check very carefully which components are used. I am very close to make a (new price) comparison with ITX vs. minipc. Of course ITX wins with DT.
The challenge is now to get the best possible cooling.
I never had expected problems with PCIe 4.0.
If I decide for the ITX, I am not building it mysel, but the dealer.
I took a quick look at powerful Mini-ITX PCs and oh boy they are expensive, new were easily over 2200 ā¬ā¦
Thatās as expensive as equivalent laptop and the cooling may not be the greatest too
The model at 2200-ish was an Ultra 9 185H, 32 GB DDR5 and RTX4070 8GB (most likely mobile, I think desktop is 12GB)ā¦ But this was only one store and I looked at dGPU models
Didnāt we see some benchmarks where it seemed that the integrated GPU of some Apple chip was very fast, probably because of a lack of memory transfers, having access to a lot of memory, and therefore no tiling?
This is because of the very large integrated memory. With 96GB you can already run some very good LLMās locally, which you canāt do unless you spend tens of thousands of euros on a dedicated GPU with that much memory.
Framework also makes an āAIā focused desktop where the benefit is exactly the same, a decent integrated GPU and tons of soldered memory.
To be fair Apple M"X" GPUs are a special case because they are more or less comparable to a lot of discrete graphics cards when it comes to silicone area/transistors they use. Couple that with low latency and bandwidth due to sharing system memory and itās the speed we see.
@linuxuser you can also go for a mini PC and then get a graphics card and couple it with oculink (Something like a Minisforum DEG1). Itās more expensive but it gives you modularity for the future.
You can get a cheaper mini pc, like a Beelink Mini PC SER5 MAX with a Ryzen 7 6800U and 32GB of RAM for 320ā¬, a PSU for 50-70ā¬, DEG1 110ā¬, and a GPU for 200-300ā¬.
I have some Beelink. The Ryzen 5 5560U is used for a mailserver and a Ryzen 7 7735HS /SER 6 Max) is the one I am writing this post.
When I compare the Ryzen 7 7735HS with my Intel B580, the Intel is 10times faster. And that is the goal for my new pc.
My problem is that I often have to be in a certain room with very very limited place. At this place I have a lot of time for darktable and other things. Why not use this situation to learn dt better?
Any combination with a minipc is useless for me, I know there are external graphic adapters.
I have 2 choices:
Buying a minipc, install linux, test dt and send it back, if I am not satisfied.
Ordering a Mini-ITX pc with my wish components, which I cannot give back. Mini-ITX would be a little bit more expensive, but uses the components which I like most. I tend to ITX. But to choose the right components is a big challenge, because the size of the components must fit in the case. You are fighting with mm.
Is it already possible to run darktable with AI? In other words does it make sense to have a lot of system memory like 128G? I donāt mean the memory of the GPU.
I run dt sometimes with 12G and then with 64G system memory. I donāt notice a lot difference when edting with standard modules, except denoise. This depends probably which modules I use.
No, I was only explaining why machines like that existed, not that they would help in darktable. Even if darktable gets AI masks, they all use very tiny models so you donāt need exorbitant amounts of memory.