Hacker News Reader

Comment by zaroth | Hacker News Reader

And the march continues ever onwards and upwards! What will we ever do with such tremendous quantities of insanely fast memory? (run Electron of course! </s>)

Only tangentially related… So nice that the chip shortages seem to have mostly worked their way out of the system.

Used server prices in particular have gotten almost ridiculously cheap lately. Just bought a 4 node EPYC server with 128 cores, 512GB DDR4, 16x NVMe slots with 15.3TB of P5500 storage populated for $2500.

Seems like almost insanity to get so much compute for that price, and yet in 5 years, as always, that machine will be considered slow and inefficient compared to the latest iteration.

It seems like progress continues unhindered by the death of Moore’s law. The hidden toiling of billions in R&D that make this possible is truly a modern marvel of capitalism.

View on HN

> yet in 5 years, as always, that machine will be considered slow and inefficient

Keep in mind it's only that cheap because someone already considers it slow and inefficient.

High energy prices definitely play a huge role in the upgrade cycle.

The higher the energy cost, the sooner that upgrading to more efficient compute pays off. But it’s still amazing to me that just the CPUs for this system were selling for $10,800 in 2020.

It's a great time to shop for CPUs. GPUs, on the other hand...

If you're planning for non-24/7 usage, and space isn't at a premium, one man's trash is another man's treasure.

What will we do? Run local LLMs. Train and fine tune LLMs. Play games full of AI NPCs. Run private local AI home assistants.

AI is the biggest driver for power at the endpoint since when 3D games came out. (I remember when Quake drove an entire generation to upgrade their PCs.)

Or you could just use ClosedAI APIs and all that will lead to…

>Just bought a 4 node EPYC server with 128 cores, 512GB DDR4, 16x NVMe slots with 15.3TB of P5500 storage populated for $2500.

That is a 32 Core EPYC CPU, 128GB DDR4, 4TB NVMe SSD per node for $625. I am assume you mean physical core and threads. But I am surprised you could get an 32 Core EPYC for $625, unless it is not even a Gen 3 EPYC but much older. The 128GB and 4TB together would have costed at least $200. Meaning you need to get the CPU for less than $350 to leave some cost with PSU, fans and Server Case.

Either it is heavily discounted or this is a 2nd hand deal.

>Seems like almost insanity to get so much compute for that price, and yet in 5 years, as always, that machine will be considered slow and inefficient compared to the latest iteration.

Likely not.

That is because DRAM and NAND price has fallen to close or below BOM cost. Your 512GB DDR4 and 15.3TB SSD is less than half if not close to a quarter of its price 18 months prior. The EPYC CPU also has stepper discount due to slower HyperScaler expansion. Basically unless we have two or more node cycle to bring down cost, the current purchase price wont last.

Can you point me in the general direction where I might be able to find some similar machines? EBay? I’m needing to build out a small scrapyard server farm and I have basically unlimited solar energy available so I don’t mind using last-gen stuff.

I keep an eye on the “Deals” forum at ServeTheHome. They link to particularly good deals on eBay.

If you don’t see specifically what you’re looking for, check out the listings and look for other items being sold by the same sellers.

You can also find sellers on eBay and often find their own storefront online.

After spending some time watching it you’ll even get a feel for what’s coming “off lease”.

Apart from eBay, which is one of the good sources, you can map those sellers to their standalone online presences, where prices are much more stable.

Companies decommission servers all the time. Look for the companies that buy them. Data protection laws, however, had made storage devices trickier to get.

May I ask where you got your machine?

If you don't need something quite that fancy, I find LabGopher is a good place to check: https://www.labgopher.com/

Reply on news.ycombinator.com