Hi all!

 

With Christmas coming I convinced my girlfriend that my loyal server (E5-2630 v3 @ Supermicro X10SRA-F) really should go into retirement and that it's time we bought something new. Still not entirely sure how, but I seem to have convinced her. This means, I now need to think about what hardware I should buy.

 

What do I use my server for?

 

I will sum it up, so it's easier to read:

 

• Traditional file share storage (samba);
• Virtual machines (win10/11; nothing fancy (ie "rdp-to-work-machine and 'download-client-machine');
• One VM which is heavier: win10 VM running Blue iris, managing 8 IP cameras around the house;
• Docker (ie Plex, tautulli, pihole, unifi controller, backup tool)
• Something I want to start doing: GenAI with Stable DIffusion on my server;
• Something I want to start doing: LLM on my server.

 

My budget:

 

Not a fixed budget really, but I don't need to end up at eur/usd 10k either :). I'd rather pay a bit more so I have something which will last longer, than saving some money and needing to buy somehting new in 2 years again. For now, lets say the budget is $/€ 4000, but feel free to convince me I need more if that's the case.

 

Hardware I am currently considering:

 

• Case: Jonsbo N5 (trying to get rid my 19" rack due to physical space)
• PSU: be quiet! Pure Power 12 M 1000W 
• Motherboard: AsRock TRX50 WS
• CPU: AMD Ryzen Threadripper 7960X
• CPU Cooler: Noctua NH-D9 TR5-SP6 4U (height fits)
• RAM: 2x kit "Team Group T-Force Xtreem FFXD532G7600HC36FDC01" (2x 2x32GB DDR5 ecc reg @ 7600)
• GPU: Gigabyte GeForce RTX 4060 Ti Windforce OC 16G
• NVME Storage: See question #3 below please.
• Spinning rust Storage: I've got some new 20TB disks..

 

Please provide feedback where you have any. Any feedback is welcome!

 

Questions I already have (may be added/updated  )

 

1 In all honesty, I am still very much doubting whether I really need an expensive Threadripper. For the past years I have been reading about the importance of ECC reg. memory, but for the past years, my normal DDR4 in my current server have also been doing it's job just fine ;).  I have been looking at the Ryzen 9950x and look forward to seeing the new Intel CPUs (to be released in a few days). Think I should stay away from Intel 13/14gen, due to the issues they have been having. (is that a good conclusion?) Also, consumer motherboards often seem to have very limited PCIe slots, looking at what I might need to have: 1 or 2 GPUs? (see question 2), 10Gbit nic, HBA(see question 4)

 

2 GPU. WHat I would like is one GPU, which I can use for booting the machine, share over multiple VMs and Dockers. I remember from the past I had to put an extra GPU in my old server so it would boot from that one and then my P2000 would be available for Plex. is this still the case? Ideally, I would buy a 4060Ti (for example) as my only GPU in the server and then use it for Plex Transcoding, Stable Diffusion (docker), LLM (docker) and also pass it through to one or multiple WIndows VMs (ie my VM for BlueIris, using AI and thus a GPU). If not, what will/will not work and would I need additional GPUs?

 

3 NVME. This AsRock motherboard has, if I am reading correctly, 1x 5.0 m2 slot and 1x 4.0 slot. I would like to put in 2 2TB 4.0 nvme sticks and use one for retention so I have 2TB, which I would use for storing docker/VM stuff. This means I am giving up the additional speed of the 5.0 slot. Should this be 'fine' as I think, or is this a silly idea as there is a much better solution available?

 

4. SATA. The motherboard contains 4x sata and 2x "SlimSAS (PCIe Gen4 x4 / 4 SATA3)". The regular sata ports I understand ;). What's up with the SlimSAS? Is this a port for which I can use a cable which splits a SlimSAS port up into 4 normal SATA ports? Would this mean I could use a total of 12 SATA disks? I will be starting with only a few disks, but being the data hoarder i am, it is likely to grow. Do I indeed want to use these ports, or do I again want a used LSI HBA card from Ebay for connecting spinning rust storage?  

 

5. NIC. This board comes with a 10gbit nic, namely the 'Marvell AQC113'. WIll this work under Unraid? Also, do I want to use this nic? I see alot of people prefering Intel cards such as the X520, but what's the benefit for me? Can/should I be ok with the Marvell, or do I need an extra card?

 

Thank you!

I understand this has been a long post, so thank you for spending your time reading everything. I tried breaking everything down as good as I can. SHould you have any questions, please feel free. SHould you have anything feedback or answers to anything of the stuff above: highly, highly appreciated!

 

Edited October 21, 20241 yr by MaTi
spelling error

40 minutes ago, MaTi said:

2 GPU. WHat I would like is one GPU, which I can use for booting the machine, share over multiple VMs and Dockers.

 

You shouldn't share a GPU with Docker and VMs.  You will have issues (crashes).  Think of it as a car with two steering wheels and two drivers.  As long as only one person is turning the steering wheel all is OK.  When two start turning?  The car doesn't go where either driver wants, and usually ends up off the road.

1 hour ago, ConnerVT said:

 

You shouldn't share a GPU with Docker and VMs.  You will have issues (crashes).  Think of it as a car with two steering wheels and two drivers.  As long as only one person is turning the steering wheel all is OK.  When two start turning?  The car doesn't go where either driver wants, and usually ends up off the road.

Thanks for your feedback! I was hoping for some magical layer between hardware and software which would manage resource and request handling I guess..  Now, to make sure I understand you correctly: do you mean a seperate GPU for each docker and a seperate one for each VM? Or: one GPU for all VM's (shared) and one GPU for all Dockers (shared). Or even: 1 GPU for all dockers shared and for each VM a dedicated GPU where needed?

 

thnx!

Epyc may be a better option on price, especially in the used market. Hard to beat when a big need is lots of pcie lanes and x16 slots.

You won't necessarily know if ECC would help unless you're monitoring for bit drift etc. in other ways. If something only lives in memory, you may just never know. And you may be fine without.

Case... Just make sure you're going to be able to move enough air for multiple gpus and everything else. I have a rackmount on a wall (bottom of the case to the wall, airflow left to right). So much easier to work in and SSD temps down 20° (previous case was an older lian li midtower)

10 hours ago, MaTi said:

Now, to make sure I understand you correctly: do you mean a seperate GPU for each docker and a seperate one for each VM? Or: one GPU for all VM's (shared) and one GPU for all Dockers (shared). Or even: 1 GPU for all dockers shared and for each VM a dedicated GPU where needed?

 

A GPU can be shared by the Unraid OS and the Docker service and the containers running that service (as long as the container has the needed resources to use the GPU).  So containers such as media servers (Plex/Jellyfin/Emby) and others (Immich/Frigate/etc) all play nicely with each other, as the Docker service manages the GPU resources.

 

With VM, things get more difficult.  It relies on using either virtualized devices or hardware which is "passed through" to the VM.  One will "Bind" the device (such as a GPU or a NIC) to the VM by selecting the IOMMU Group the device is in, through System Devices in the Unraid Tools tab.  Once the device is bound, it is no longer seen by the Unraid OS, so it cannot be used for Docker containers (or for Unraid GUI access).  Binding the device allows you to select it in the VM template for use in a VM.  You can assign the device to more than one VM, but you would only want to have one of those VMs running at the same time to prevent conflicts.