Plenty of computer for the price. Have bought from this offshoot of ACT before and had a positive experience. CPU isn't the most efficient. GPU can game at about the same level as a gtx1650. The 500w power supply has a 6 pin connector, verified from Lenovo documentation, so should be able to drop in a more performant card if desired, evidence
This would still be pretty decent for a plex machine. For 300 bucks and 2TB of storage and ability to decode… you could do a lot worse!!
You could get a SFF with a brand new HDD for the same price though, it'd use less power too.
SFF won't have a GPU, and a secondhand low profile Quadro card is relatively expensive.
@coxymla: Unless you're trying to serve 20 users or something, it shouldn't make a difference for Plex.
I wouldn’t be trusting those HDDs
If it was going to fail it probably would have by now. MTTFs are so high these days. But anyone using it for any storage would ditch that and stick in multiple higher-capacity ones.
What's the card to have now?
To save you an Ark this is a 4C8T skylake CPU.
CPU was launched in 2017 and has a TPD of 120W
Is this a good device for an AI Servers? Ollama or deepseek etc?
No
rip thanks
What do you need for that? High RAM or just a beefy Nvidia GPU?
Kinda both. You need to have capacity to load model (iow enough RAM or GPU RAM) and you need compute capability to run inference processing. There are different versions of the same model with different quantization levels. Quantization reduces the size of an AI model by lowering the precision of its numerical values, typically converting 32-bit floating point weights and activations to 16-bit, 8-bit, or even down to 1-bit representations. Smaller quantization reduces the model size and makes inference faster, but also reduces accuracy (which may lead to wrong answers, hallucinations etc…).
When you see somebody saying they loaded something like Deep Seek on Raspberry Pi that means is one of those super low quantized models (possibly even cut down in other ways) and it is done for clout and view harvesting.
There are some interesting projects that allow you to run distributed inference across heterogenous devices where the size of the model you can load is limited by total available RAM size across whole device pool. If you are interested in that check
There is Twitter/X post how to run Deep Seek at home without GPU for US$6000 which is considered cheap
@dr: Good lord, mad respect for putting in effort to explain things to internet randoms. Also love that the effort of the responses was proportionate to the relative effort of the questions.
@dr: Some more details about that configuration in here: https://www.pcgamer.com/hardware/graphics-cards/today-i-lear…
When I asked about big amounts of RAM I did not expect 768GB of them.
This seems to have 768GB of RAM for a little bit over $1,500:
https://www.ebay.com.au/itm/186855905323
When you see somebody saying they loaded something like Deep Seek on Raspberry Pi that means is one of those super low quantized models
Not only that, it will invariably be one of the "distilled" R1 models - which isn't really R1 at all, but other base models like Llama 3 8B trained on R1 outputs.
The actual R1 is so huge that even when highly quantised it's still hundreds of gigabytes in size, well beyond what most systems can cope with.
Thanks for letting me know. Seems its out of my reach at the moment.
Nice ECC ram. Good budget nas box
Electricity costs though would be high
Yes, but the hardware is cheap. Would be 1k if you buy a synology/qnap
Definitely. I’m running an old Core 2 Duo machine with 16GB as a gateway server/firewall with FreeBSD. Half keen to get a nice Xeon with a bit more RAM and ECC to boot. But I’ve been switching to NVMEs for my ZFS drives, I guess a board of this vintage wouldn’t have any NVME slots (currently using a PCIe expander).
Heyo, I think it has 2 NVME slots, one used, one spare
WIll this be any decent for gaming like GTA 5?
The GPU is a cut-down GTX 1060. It will be OK for GTAV, probably around 60 FPS on high (not ultra) at 1080p.
If you literally only have $300 and can't possibly spend any more, this machine is probably the best thing you can get for the price.
Most people would probably prefer to spend around $500 on a second-hand gaming PC.
What would you get for $500 as 2nd hand gaming PC?
For $500 I reckon a good deal would be a Coffee Lake or early Ryzen 6 core CPU, and a strong Pascal/RDNA or a middle Turing/RDNA2 GPU.
Would also like to know
I have a P520 (non-c) for dev/ML use - it's a better option if you want to add power hungry components, as it can be specced with up to a 900w platinum PSU.
Can be found for not much more cash if you're willing to watch ebay listings.
I'm curious about the hardware specs of your machine? Can you share them please? I presumed you beefed up the RAM and put GPU (or few) there
Yes I did.
- Xeon W 2135
- 64gb ddr4 2666mts quad channel
- AMD MI100 32GB
The base system was $350, memory and nvme upgrades about $150, GPU $1050. The MI100 is very beefy silicon but usually not as well optimised as nvidia, so it generally performs like a 3090 with more vram.
Nice!!! Thanks!
Got one! Thanks.
Please note that this one's NIC is pretty lame, just a basic 100Mbs speed card and you need to buy a new one for better speed in case of building a home server/NAS
![[eBay Plus, Refurb] Lenovo P520c: Xeon W-2123, 32GB RAM, 256GB NVMe, 2x 1TB HDD, P2000, 500W $300 Delivered @ ACT Networks eBay](https://files.ozbargain.com.au/n/19/891819.jpg?h=db56052e)
Not that long ago the p2000 was the card to have for plex transcoding.