Is this a good deal for a server to use as a NAS?

[is39]

Well, there is some logic behind this: while SAS SSDs are generally twice as fast - not a hard rule, but most of SAS SSDs I've seen are SAS12 or 12 Gbps vs 6 Gbps on SATA SSDs, but on the other hand, Enterprises generally don't buy used hardware, and for consumers, SAS SSDs are harder to use since requires specially controllers, generally more power usage, and most importantly - require active cooling as they get quite hot.

I agree on the causes; but those SAS SSDs did cost more money when sold (vs SATA counterparts of similar capacity) and are faster, so one would expect them to continue to hold value. I guess since everybody can use SATA SSDs this caused scarcity on secondary market and driven the prices up (vs SAS or NVMe SSDs). Now with AI devouring primary market things can only go worse, until that AI bust everybody is anticipating ;-) :-(

--igor

 

[TrevorH]

Given the listed bits of hardware in the first post with a price of £850, I picked the absolute cheapest parts that match these from ebay UK and just breaking it up into bits and selling them makes a huge profit. I was mean with the prices too so most of the 2.4TB drives go for more than £100 but there was one that sold for £80 and I used that.

8 x 2.4TB SAS drives @ £80, £640
4 x 4TB SAS SSDs @ £200, £800
4x32GB DDR4 2666MHZ RAM, £350
HP Proliant DL360 Gen10 with 2xXeon 5118, £220

Also I saw mention that lots of storage in one drive would be preferable to lots of smaller drives but that really depends on your purpose. If you want to run an active database then having 8 spindles that can all seek at once is definitely a good thing.

 

[stonewall]

It takes years to recoup the cost of lower energy use hardware at my usual consumption, about 8 hours per day, five days a week. These are the 8 3.5" capable tower housings I thought would work for me. The hardware prices are what I saw recently. If you have to buy memory for the low energy options the prices are insane. I already have the RAM for the energy guzzlers.

Upfront Cost vs. Long-Term Energy Expense

Model	Upfront Cost (Excl. VAT)	Est. Annual Energy Cost (Light Load)	Total Cost After 5 Years (Light Load)	Est. Annual Energy Cost (Full Load)	Total Cost After 5 Years (Full Load)
Dell PowerEdge T430	£149	£62.40	£461	£91.52	£606.60
HPE ProLiant ML110 Gen9	£200 (+£51)	£39.52	£397.60 (-£63.40 vs. T430)	£70.72	£553.60 (-£53 vs. T430)
Dell PowerEdge T330	£200 (+£51)	£33.28	£366.40 (-£94.60 vs. T430)	£59.49	£497.45 (-£109.15 vs. T430)
Dell PowerEdge T340	£425 (+£276)	£17.47	£512.35 (+£51.35 vs. T430)	£59.49	£722.45 (+£115.85 vs. T430)
HPE ProLiant ML110 Gen10	£425 (+£276)	£18.72	£518.60 (+£57.60 vs. T430)	£38.27	£616.35 (+£9.75 vs. T430)

Given the listed bits of hardware in the first post with a price of £850, I picked the absolute cheapest parts that match these from ebay UK and just breaking it up into bits and selling them makes a huge profit. I was mean with the prices too so most of the 2.4TB drives go for more than £100 but there was one that sold for £80 and I used that.

8 x 2.4TB SAS drives @ £80, £640
4 x 4TB SAS SSDs @ £200, £800
4x32GB DDR4 2666MHZ RAM, £350
HP Proliant DL360 Gen10 with 2xXeon 5118, £220

There are just 2 x 3.8tb SAS-3 HP enterprise SSD drives, with the other things you mention. But it would be a good profit to sell everything.

 

[is39]

NAS typically runs 24/7; while it's possible to spin drives up and down, with transition to SSDs the savings are less;
there is also argument that repeated spin-ups/spin-downs kill hard drives (i've a few anecdotal proofs of that).
Plus nowadays NAS often runs other services (virtualization, logging etc), which means it cannot be turned off or standby-ed.

Even with modest 100W power consumption and assuming $0.30 per kWh (which is low in PG&E land, and run on par to UK average 25 pence) we see 0.1 kWh * 0.30 $/kWh * 24 h * 365 days = $262 per year, which is sizeable budget to optimize power costs. My NAS-es usually run for ~10 years or so, but it's an outlier i suspect; still, even with very reasonable 6 years there seems to be space to save on lower consuming components. Things look quite differently if you've cheaper power ($0.18 kWh which i pay) or more expensive hardware (like what we see now). But with AI competition for electricity i expect our power prices to keep increasing; which means solar and lower-powered/power efficient hardware is good investment.

--igor

 

[stonewall]

I looked into buying a Rx40 rack mounted server and doing a case swap to an XL case. But the PSUs for these servers basically plug into the motherboard. It would be a hell of a job to move them somewhere else, and solder in cables to the motherboard. There are more headaches moving the hard drives.

It turns out my HP Z440 Nas has 64gb DDR4 Ram in 4 x 16gb ram modules (MTA36ASF2G72PZ-2G1 - Micron 16GB DDR4-2133 RDIMM PC4-17000P-R Dual Rank x4 Module). I'd forgotten I had this much. Ram was so cheap a couple of years ago it wasn't worth consideration.

Realistically, I am not going to move to a system where I can't use that memory unless I get an amazing deal on a Dell T340 or HPE ML110 gen10 and I can afford the ram needed for that system. I think this memory works in the HPE ML110 gen 9. I will probably go for a case swap, to give me more drive space. People have done this quite a bit with the HP Z440. They buy a new high quality PSU, for the longer cables, as the HP Z440 cables are restrictive, and a new/old case.

Even with modest 100W power consumption and assuming $0.30 per kWh (which is low in PG&E land, and run on par to UK average 25 pence) we see 0.1 kWh * 0.30 $/kWh * 24 h * 365 days = $262 per year, which is sizeable budget to optimize power costs. My NAS-es usually run for ~10 years or so, but it's an outlier i suspect; still, even with very reasonable 6 years there seems to be space to save on lower consuming components. Things look quite differently if you've cheaper power ($0.18 kWh which i pay) or more expensive hardware (like what we see now). But with AI competition for electricity i expect our power prices to keep increasing; which means solar and lower-powered/power efficient hardware is good investment.

Price the parts off ebay for the UK now. The power efficient systems, like T340, ML110 gen10, are expensive, even with 8/16gb Ram, then add the extra ram cost on top (very expensive), and budget for larger hard drives, say 24tb, or SSDs, to get the power use down. You might save £100-200 per year and have the deal pay off in 3-6 years. I think AI will cause a lot of power efficient hardware to dump into the market in the next to 6-12 months. Probably best for me to struggle on and wait. I can do my case swap and PSU buy for £100-120.

 

[is39]

I've already noted that when/if we'd ever get power efficient hardware from AI (debacle?), in many cases it would be too powerful for the lab user, or require a lot if supporting infra (rack-sized central power, liquid cooling). Obviously some datacenters would be dumping previous gen hardware, some if it still with air cooling etc (though most AI datacenters are new builds, so there is nothing to dump). At least NAS-es are NAS-es, sort of, so those older may get reused (though 400GbE and even 100GbE networking is too much for me, for now).

 

[MiniKnight]

I've already noted that when/if we'd ever get power efficient hardware from AI (debacle?), in many cases it would be too powerful for the lab user, or require a lot if supporting infra (rack-sized central power, liquid cooling). Obviously some datacenters would be dumping previous gen hardware, some if it still with air cooling etc (though most AI datacenters are new builds, so there is nothing to dump). At least NAS-es are NAS-es, sort of, so those older may get reused (though 400GbE and even 100GbE networking is too much for me, for now).

2 problems:

1. You may not like it, but it isn't a debacle. AI compute is growing fast. At our company we're at 70x tokens/ week compared with October
2. It'll be like used servers today. Xeon E5's are cheap but anything before 2nd Gen Xeon isn't usable b/c of security. So you can get old chips that nobody wants and can use cheap, but they're getting out-performed by Panther Lake 2:1 at a fraction of the power. V100's are cheap, but they're slower and you're paying power. Today's GB300 station will be old in 5-7 years and cheap, but that doesn't mean it'll be good.

I don't get bag over head and hope AI will pass. If you work at almost any business you're seeing token use go up fast.

I'm not saying AI is always correct, and that it isn't a security nightmare. What I'm saying is that it's happening fast that I wouldn't expect prices to go down soon.

 

[is39]

Well, there was a question mark behind that debacle above ;-)
I use AI daily, and it's quite useful, even with it's limitations. It does ruin many jobs, but it's a natural (sort of) order of things.
It's also overhyped, like Internet was. Or not, in which case AGI would make us all into paperclips ;-)

All I've meant is that it's unlikely we would be able to use most of AI datacenter hardware at home labs, due to power, cooling and other limitations.
Smaller scale units, like GB10 based boxes, or apple silicon stuff, or even regular PCIe GPUs - sure; but those would not come from AI datacenters (well, i guess those V100s, or better A40s etc would come eventually).

Now, I could be wrong. I do use enterprise switches and 1U servers. Power costs go down with solar.
But parting up GB300 rack... I just do not see it yet ;-)

 

[stonewall]

I've already noted that when/if we'd ever get power efficient hardware from AI (debacle?), in many cases it would be too powerful for the lab user, or require a lot if supporting infra (rack-sized central power, liquid cooling). Obviously some datacenters would be dumping previous gen hardware, some if it still with air cooling etc (though most AI datacenters are new builds, so there is nothing to dump). At least NAS-es are NAS-es, sort of, so those older may get reused (though 400GbE and even 100GbE networking is too much for me, for now).

I was thinking of companies ditching regular hardware to upgrade because it does not support AI. That could be a lot of interesting hardware for us.