Home lab for virtulisation, containers, AI and ML

[BWMerlin]

I am looking to finally put together a dedicated home lab for self study in managing virtual infrastructure, containers, AI and ML as well as perform general home server duties (maybe even a forbidden router).

I was thinking about building around the Intel W LGA4677 series in a Thermaltake Core W200 Super Tower Chassis for a two in one case for future workstation/gaming PC build.

So far the only components I actually have and would like to use are:
6 x Intel Optane 905P Series 960GB U.2
2 x Intel SSDPEK1A118GA01 Optane P1600X 118 GB Solid State Drive - M.2 2280

I was looking to run Proxmox as the hypervisor on a mirrored array on the two Optane P1600X M.2 drives and then run a tiered storage pool with the Optane 905P drivers and some cheap flash storage (would like to stay away from the spinning rust).

What I am looking to actually do on the server is labbing Windows domains (simple domains, inter domain trusts and Intune/SCCM and other hydration kids so the ability to do nested VM’s is critical. On the Linux side of things I will be looking at working with Ansible to do IaC as well as labbing integrating Linux into a Windows environment to build skills around what might be typically found in a modern network.

For the container side of things I have never touched containers before so I am looking to go from setting up a simple container to run a media server or the like to then progress into full container orchestration for scaling etc.

As for AI and ML I don’t even really know where to start with these two so the ability to start from zero with a “good enough to get you going” setup is all I really need for this.

What I have come up with so far is this
Intel XEON W5-2445
ASRock W790 WS
4x Kingston 64GB ECC Registered DDR5 4800 (PC5 38400) Memory
Thermaltake Core W200
Noctua NH-U14S DX-4677
FSP Twins Pro ATX PS2 1+1 Dual Module 900W 80 PLUS GOLD

Plus the following which I already have
6 x Intel Optane 905P Series 960GB U.2
2 x Intel SSDPEK1A118GA01 Optane P1600X 118 GB Solid State Drive - M.2 2280

And looking at either an AMD Radeon Instinct MI25 or MI50 from ebay

None of the above (other than the Optane which I already own) is set in stone so any feedback on suggested alternatives is appreciated.

[BWMerlin]

@wendell would love to hear your thoughts as to if this is going to be a viable build or if I should consider some changes?

[wendell]

might be able to loan you a tesla v100 to get you started if you want. unless you go for 328452 cores you dont need a lot of cores to feed the gpu. its a solid 10 year system you’re building

tho you might be able to get a better deal on closeout cascade lake or skylake xeons, or hell even epyc rome is shockingly cheap on ebay. its not as nice as the w790 platform but for ai/ml you probasbly arent doing much on cpu anyway.

you might be able to get a rome barebones with u.2 bays cheap which makes a lot of the connectivity cost go away and is cheaper than diying it with the atx psu etc.

mi25/50 is too EOL imho

MI210 or castoff MI250 chassis would also be ideal, except that the software stack on those has improved so much anyone with those is not letting go… even though they are a lot slower than MI300

[BWMerlin]

@wendell

might be able to loan you a tesla v100 to get you started if you want.

Thanks for the generous offer, once I actually get this thing built I might reach out.

@wendell

you might be able to get a rome barebones with u.2 bays cheap which makes a lot of the connectivity cost go away and is cheaper than diying it with the atx psu etc.

Is that because of the need of a HBA vs direct connection to the motherboard?

How sensitive are workstations/servers to none kit RAM? There is cost savings to buying multiple single RDIMMS vs buying a kit of the same capacity and was wondering if it was worth the extra cost to buy RAM kits vs buying the single RDIMMS.