Multi-million dollar Cheyenne supercomputer auction ends with $480,085 bid | Hacker News Reader

Multi-million dollar Cheyenne supercomputer auction ends with $480,085 bid

tomshardware.com

55 pointsdangle135 comments

> Still, a savvy eBay seller could flip the processors and RAM across the machines for around $700,000 (£550,000), making a hefty profit.

> At a conservative estimate of $25 million for the array to be constructed, the auction sale is only 2% of the sticker price of Cheyenne.

Honestly $480k sounds like a fair price. This is 8 year old hardware, in any reasonable depreciation scheme the book value is $0. If you just flip it you probably break about even, considering that prices are not going to be able to sell 8000 of those Xeon processors at their current market price.

If you just want the compute, this supercomputer is equivalent to about $70000 in modern GPUs (assuming your compute needs fit the processing model of GPUs), so even with free power it's not terribly attractive to buy the system to run it yourself.

You can’t flip the CPUs instantly, but if you’re a company that supplies older parts to people who need them, and can stock them on the shelf, you ought to be able to get most of the market value - eventually.

I’ve paid for a motherboard of a off-lease hardware because even though it was about the price of a tune new off-lease server, it was simpler for me than rebuilding the OS for a newer piece of hardware. And data center equipment is so fast to swap even motherboards. If a CPU burned out in this I’d just get another.

Yeah. There was a time back in like 2016-17 where i think it was Facebook that was decommissioning an old data center and that alone crashed the market for E5 v2 Xeons

Which raises the question, is there a blog | twitter account tracking datacentre decommissions | revamps with related hardware expected to be released to second market for people that want good deals on X year old then bleeding edge gear?

It's extremely regional.

It would be better to invest on something what would track the things you need and notify you if the price drops or there is a surge in availability.

And an anecdote to complement this:

I have 1155 desktop PC with Pentium G2020 (2cores/2threads, ugh). I can fit almost anything Ivy/Sandy Bridge in the mobo, but...

i7 3770 is still $300, while various refurb Xeons E3 v2 are $20-30.

And the best part is what even if I buy something like 1270v2 (4c/8t, full eq for 3770) it would still be a more than a decade old system and it wouldn't improve the performance much.

It’s exceptionally regional once it is full systems as they’re heavy as hell. Find a local scrapper and get in touch and you can often save tons by avoiding shipping.

Years ago i went and looked for most powerful 1155 processor my old motherboard supported and i think it was i7 3770 - i bought it for 20usd. How it can be 300 now?

They’re not. I instantly see them for $50 on Amazon and $34.99 on eBay.

I gave the price for a new one from the retailers (they somehow still have 1 or 2 of 1155 in stock).

There are, of course refurb/secondhand 3770 out there, but...

Just let me copy-paste one of the lots out there:

    $33 e3 1270 v2 SR0P6 3.5-3.9mhz tdp 69w = core i7 3770k
    $39 e3 1275 v2 SR0PA 3.5-3.9mhz tdp 77w = core i7 3770k with Intel HD Graphics P4000
    $44 core i7 3770 SR0PK 3.4-3.9mhz tdp 77w
    $43 core i7 3770S SR0РN 3.1-3.9mhz tdр 65w

    I have 1155 desktop PC with Pentium G2020 ... even if I buy something like 1270v2 ...
    it would still be a more than a decade old system and it wouldn't improve the
    performance much.

That depends. For the small 3 node cluster of HP Microservers (Gen 8) in my home lab, they had cpus of the Pentium G2020 class. Upgrading those to the Xeon E3-1265L v2 (also 4c/8t) made a huge difference.

The G2020 has 3MB of cache, whereas the E3 above has 8MB and clocks higher.

Then again, the workload on these little boxes is mostly running VMs for build jobs. More cores is definitely better than less cores for that. :)

I’m pretty sure you can’t just up and toss it on eBay like the article says, the supercomputer is subject to export controls.

$25 million also seems likely to have included labor, cooling, and possibly software and support, none of which are saleable capital assets.

> Still, a savvy eBay seller could flip the processors and RAM across the machines for around $700,000

Three (pretty big) caveats to add:

* Any significant quantity of those CPUs and RAM sticks hitting the used market at once will depress the selling price. So they need to be sold over several years during which the baseline demand (and price) will continue to fall.

* The sale contract had fairly onerous requirements about the professional personnel and methods of removal permitted, all of which will add to the buyer's upfront costs on top of transport, storage and sales costs.

* Since there are known failed components in the lot, all the pulled parts will need to be tested or sold as-is.

I'm surprised it got bids that high and I think the article implying the buyer got a steal is overblown.

First caveat is even bigger. There already is a HUGE amount of those Xeons and ram modules on the market coming from China where old servers get scrapped, and market is not soaking them in. Nobody needs $18 ancient Xeons. You will never sell 10K of those CPUs in your life time, even if you lower price to $1 a pop.

Wonder how much gold (etc) is in these chips then?

Might be a case of actual recycling being the best course of action.

Buying this supercomputer for 480k and parting it out for 700k doesn't sound profitable. The amount of labor involved in that sort of business is substantial, you have to store the whole thing while you do it, and you're going to crash the ebay market for those parts to boot. Unless the buyer already has buyers lined up for bulk purchases, I don't think parting it out would make any sense.

If they bought the super computer with the goal of doing actual computation I really hope they know what they're getting themselves into. For example: if you just look at the clock speed and the cores you might think that you can calculate some kind of cost to do the work you need. But that would be extremely simplistic because the way CPUs work is so complex. CPUs have features that will execute certain work loads faster, that can cache certain instructions, and be more efficient with the same resources. So you might have CPUs that have a lower clock speed and core number that still blow older CPUs (that 'look better') out of the water.

It can be hard to know how well your CPU is going to perform with the instructions you generate until you have it tbh. You could get a sense of how fast it is compared to other processors with benchmarks. But I don't know how accurate it's going to be overall. A lot of the time you end up with unexpected results; CPUs that seem cheap are usually cheap for a reason though.

I mean, this is literally what linpack was invented to do. It is used as a standard benchmarking tool.

You can get these chips on eBay for quite cheap now. I run a couple in a dual Xeon home server.

But imo, the power consumption is just not worth it. Might just switch to a regular desktop running an i5.

Could you use the hardware to start a Low end cloud provider? With pricing per core many low end users won’t really know the difference…

I kind of wonder if the profitability of these types of companies are based on perf/watt, or more directly electricity cost.

Sort of how the electricity of mining a bitcoin is orders of magnitude more expensive than just buying a bitcoin.

Yeah, cloud computing is a streightforward energy efficiency business. That's why the pricing of enterprise hardware is always pretty frontloaded.

It’s certainly not orders of magnitudes higher.

A S21 Hydro can do a terahash for 16 joules.

At the current difficulty, you need approx 525,000 kWh to mine a bitcoin.

At $0.16 a kWh, you’re looking at $84k to mine a bitcoin, worth $63k. That’s not orders of magnitudes higher.

If you can get cheaper electricity and are willing to pay the capex then you can make a profit.

wow, thank you for doing the math.

I guess I haven't been keeping up with bitcoin vs electricity cost. I was going to buy a bitcoin miner when bitcoin was ~$500 and in short order the electricity cost of running the miner exceeded the cost of bitcoin and started getting worse.

Now with bitcoin value > 1000x ... hmmm. maybe my assumptions are plain wrong.

Looks for corresponding auction bids on any solar panel farms getting liquidated? :)

I would think so although at the very low end I could see it being a per core post instead of compute.

Unclear how much of the cost is hardware vs energy vs infrastructure. I would assume hardware is the cheapest part but I could be wrong

See the previous thread - I called it ;)

https://news.ycombinator.com/item?id=40201184

In the spirit of Dang, because he's AFK or something this rainy day:

Thanks! Macro-expanded:

Multi-million dollar Cheyenne supercomputer auction

https://news.ycombinator.com/item?id=40197277 (5 days ago, 154 comments)

P.s. bombcar, would you be so kind as to drop me an email at metadat.luvs.u@gmail.com ? I would like to keep in touch, you're amazing.

They bid too high

Why was this decommissioned?

Don't know for sure, but from the article:

> This high failure rate is the reason for the sale, with upkeep made challenging by the fact that Cheyenne "is currently experiencing maintenance limitations due to faulty quick disconnects causing water spray."

I don't believe that.

Why not? Pipes and connections can leak? If all were installed at the same time the same way if one leaks likely others will too.

Darn, and I was just about to offer $480,086.

Reply on news.ycombinator.com