Refurbed Thinkstation 920 - Spec'ing and upgrading for machine learning

Hi Everyone,

I'm hoping someone can help me?

I've never built a PC before, and don't have the time to go all the way down the rabbit hole learning how to build one from the ground up (as much as I'd like to). With that in mind, I've been looking at pre-built systems that allow you to pick the configuration (i.e. "getting experts to build it for me").

I currently have a MacBook but I need a machine that I can run quite complex simulations on, perform machine learning with and locally run some LLMs.

As budget is only ~USD 2,500 (I say "only" as it doesn't seem to go that far, these days) I'm looking at older, refurbished hardware vendors as a way to buy a decent machine, and in particular looking at the Thinkstation 920 as a) I've had Lenovo Thinkpads in the past and always really liked their quality, 2) it appears to be built like a tank, and 3) it was $$$ when new and sells for a lot less now.

My questions about building for machine learning, specifically:

1. The vendor has a variety of Intel processors with names that mean very little to me. They all seem to be "Intel Xeon [Gold or Platinum] [5XXX / 6XXX]" but beyond the number of cores (and price), I lack a way to meaningfully benchmark one against the other. Are this era of processors any good? Should I leave the processor off the build and get a more moden processor that I can fit myself?
2. I've read online that I need 24Gb+ of GPU memory to run LLMs locally. The options available from the vendor are a Nvidia Quadro M6000 (I can afford x2 of these, if needed) or a Nvidia Quadro-RTX 5000. The performance of the latter is "better" (according to the benchmarking websites) - should I be optimising purely for memory (i.e. optimise for "Gb / USD") or for pixel rates / something else?

About the Thinkstation 920 in particular:

1. Can anyone suggest a good source for checking what GPUs are compatible with the Thinkstation 920 motherboard? (I have already tried to Google this ) I have a Nvidia GeForce GTX 1080i that I can pull out of another machine and either use instead of the GPUs mentioned above, or combine with one of them (can you mix different GPUs in the same build? Will one bottleneck the other?)
2. More broadly, is buying (slightly) older, refurbished tech a false economy? Hardware moves pretty quickly and I'm wary of commiting to a platform that will be outdated in ~18 months. Am I just being dazzled by the fact that in 2017 you could spend USD 50k+ on one of these?

Appreciate that the above is a lot, thanks in advance for any help anyone is able to provide.

R

Hi, Welcome to BC.

I am going to say spending 2.500 on a think station is not a good use of your money.    You are way off the mark.  Or You are the Mark.  

Please take a look at what you can get out of Colorado if in the US : https://xoticpc.com/      (Note make nothing from the URL, bought a laptop, shipped to CA and paid duty and shipping and no warranty to enjoy a custom LT)

Best to look further as what your thinking is dated as well maybe over kill for "perform machine learning with and locally run some LLMs."

Nvidia GeForce GTX 1080i is old. They are selling in the 4060 range of cards now.

Quadro M6000 was a professional graphics card by NVIDIA, launched on March 21st, 2015.    Dated, why waste money on over priced GPU.

Anyway, think need to look further.

Edited by Pkshadow, 24 September 2024 - 03:38 AM.

Thanks for the fast response, and for the welcome!

Quadro M6000 was a professional graphics card by NVIDIA, launched on March 21st, 2015.    Dated, why waste money on over priced GPU.

 

The M6000's cost ~EUR 400 each from the vendor, I guess whether that is overpriced is whether the "variable-to-optimise-for / USD" value is good or not. My naive, based-on-trying-to-drink-from-the-internet-firehose research was that GPU memory was the "thing to optimise for" when running / training LLMs locally (and so EUR 800 for 48Gb would seem like good value vs. EUR 1k for a 16Gb 4080 card). If I need to optimise for another variable for this type of work, please let me know . 

 I am going to say spending 2.500 on a think station is not a good use of your money

 

I should point out that was when I configured the machine with 256Gb of RAM, huge drives, x2 of the M6000 GPUs, and (almost) the highest-core processor I could see available (USD 2,500 is also a limit, not a target). If I remove the GPUs from the build and buy a more modern one new (and trim some of the other specs) then the base Workstation would be ~USD 1,200, plus whatever the GPU costs. 

Edited by hamluis, 24 September 2024 - 06:39 AM.

Ok, well the problem lays with the Motherboard of the Thinkstation.  How do you know what is in there intel what. ? and what that Boards specs are here :

https://www.google.com/search?q=Thinkstation+920+specs    And it is a Duel Chip Board.  

Would think the upkeep of drivers/software would be the problem as well.  (I have a ASUS Rampage from 2008 that still runs great but can not do extreme modern things)(there was a dual chip Rampage board but can not remember info on it).

1 company think your quoting from : https://pcserverandparts.com/lenovo-thinkstation-p920-tower-2x-intel-xeon-gold-6154-3-00-ghz-18c-256gb-ddr4-1x-1tb-ssd-1x-2tb-ssd-nvidia-quadro-p4000-8gb-gddr5-windows-11-professional-refurbished/

https://www.intel.com/content/www/us/en/products/details/processors/xeon/scalable/gold/products.html    Note Ram Speed and Ram size Compatibility.

Note : Retail price of 1 chip starting @ retail 3600.00 US.    Just for 1st on the list 6xxx.    Any whooooo I have no further info to share. 

This type of product has not been discussed in this forum that I know of, maybe someone will have more info/ideas for you.  G;luck