All posts by kobaan

Hardware, TrueNAS

Replacing my DIY NAS with an Ugreen NASync DXP 8880 Plus

Yesterday my NAS replacement hardware arrived.

Previously I was using an Intel Avoton C2750 board which failed years later with a written to death eeprom.
Then I replaced it with an Intel Atom C3558 board where it was promised that that should not happen… I wont wait…

There was a good offer, and I wanted to migrate to SSDs anyway.
The Ugreen NASync DXP 8880 plus, has 8 SATA bays for 3.5/2.5 disks,
as well as 3 onboard m.2 NVMe slots and an PCIe extension slot where an dual NVMe expander card can be installed. CPU is an i5 1235U with 10 cores.

As soon as the memory extension and extender card arrives, I‘ll post some more pictures.

Replaced the fans with some silent Noctua NF-A12x25, and replaced the RAM with 2x32GB Crucial DDR5 4800MHz SODIMM 1.1V

Now I’m still waiting for my SSD drives to be delivered.
8x Western Digital Red NAS SSD SA500 4TB and 4x WD SN850X 8TB NVMe’s.

For now I keep the 128GB boot NVMe, but it could be replaced too and hold another pool, we’ll see.

Hardware

2x 8TB RAID1 on MacOS via Thunderbolt 4

Since I’m putting huge loads on my MacStudios, I wanted to get the read/write load away from the internal SSDs and get a bit more redundancy in case an SSD fails.

All my containers running in Podman are now placed onto a SSD RAID1 on two Western Digital SN850X 8TB NVMe m.2 SSDs which I placed into an Anyoyo dual NVMe fan Thunderbolt-4 casing.

I’m still not sure if I should put all LLM model caching directories onto this RAID1 and export it from there, or put it on the main NAS and mount it from there via 10GB/s. Also MacStudio/Link via 80GB/s TB5 would be an option, but currently I don’t wont ip forwarding there…

Software, Wordpress

Blog site moved

I finally migrated my wordpress blog over from https://teddy.spacetechnology.net to https://myblog.kobaan.nl

The whole thing now runs in a podman pod on the “smaller” 256GB MacStudio, as compared to the tight FreeBSD jail before.

All my docker images are automatically updated using watchtower, and wordpress also uses auto update.

Unfortunately the migration broke 5 times, as I had to readjust resources a lot. And finally changing the domain broke 1400 links, which I had to repair. Now still some images are mysteriously unlinked although the media browser can see them. I try to fix that the next days…

LLM

LLM – Optical problem solving now works

Finally got the PDF OCR flow working, and the LLM can now solve math problems. Only grey diagrams are still hard for it to recognize properly.

I used a math training exam that my son is currently working on in preparation for Gymnasium.

And even the 27B model is sufficient to solve this level of math.

Apple, Hardware

Yet another Mac Studio M3 Ultra this time with 512GB

Did I say enough…..
meh….

Okay lesson learned, never enough.
My “1 month old” M3 Ultra 256GB went out of memory running all those models and podman containers in parallel.



My current setup is:
OpenWebUI:
-> LM Studio, Aya-Vision-32b,
-> ComfyUI workflow with t5xxl, Flux1.dev, llama-3.1
-> MCP proxified: Searxng, wikipedia, docling, context7, time, memory, weather, sequential-thinking
Podman: 24 containers including supabase, wikijs, watchtower…

Also I discovered that I can use OpenWebUI, SwarmUI, exo and even mlx
to distribute workload across both Mac Studios via 80GBs thunderbolt 5 bridging.

And with the orange clown, you never know if there will be a new M4 Ultra next year at all.

LLM

LLM full throttle on M3 Ultra 84 Watts !

What a crazy efficiency monster.
Running a LLM on all 80 GPU cores and the system is only drawing 84 Watts…..

Nvidia must be crying at night.

and this is only M3, M4 is even more efficient, but yet not available as an ultra fusion variant, and end of year Apple will manufacture in 2nm