For twelve years the garage had a sound. Three HP ProLiant DL360 G7 servers, one rack unit tall, dual processors, a wall of small fans, bought used and run hard. They were the virtualization platform for this house and, for a long time, for my clients. First as a VMware cluster. Then, last year, as a Proxmox cluster under the same metal. This month the last of them came out. The work moved onto a pair of Dell OptiPlex 7090 small-form-factor desktops. The difference is not a refresh. It is a different decade of computer, asked to do a smaller job, and the garage is quieter and cooler because of it.

The cluster is still named for ECT Consulting. That name is the whole story of why the old machines made sense, and why they stopped. Those servers existed to hold a lot of virtual machines that were not mine: customer systems, dedicated guests, the sort of density you buy a dual-socket Xeon for. That estate has been retiring, guest by guest, as the consulting work moved on. What is left is the house. Mail. Monitoring. The internal sites. A log server. The music server. The public websites, including this one. The kids’ Minecraft worlds. Four ham radio reflectors. Near the end, one of the DL360s was using about 13 gigabytes of the 192 it had installed. The rest of that memory was just warm.

What Twelve Years of Server Looked Like

A DL360 G7 is a 2010 design. The processors in mine were a little newer than the chassis, and still old: two Intel Xeon X5675s per server, Westmere, 32 nanometers, launched in 2011. Six cores and twelve threads each, 3.06 GHz, a turbo that stopped at 3.46, 95 watts apiece. Two of them is twelve cores, twenty-four threads, and 190 watts of processor before a disk spun up. Three servers was thirty-six cores and seventy-two threads, which sounds like a lot until you remember what those cores were. They are slow by the standard of a desktop chip from 2020, and they were never going to get faster. They just kept drawing power.

Around the processors sat the rest of a real enterprise server. Each machine held 192 gigabytes of DDR3, twelve sticks of 16 gigabytes, ECC, registered, 1333 megatransfers. Error-correcting memory is the right kind of memory when a box runs other people’s machines for years. It is also twelve small heaters. Storage was seven 600 gigabyte SAS drives at 10,000 rpm. Six of them were a RAID 50 behind an HP Smart Array P410i, about 2.2 terabytes the operating system could see, with a battery sitting on the controller so a power cut would not lose the cache. The seventh drive was a dedicated spare, installed and spinning so a failure would have something to rebuild onto. Across three servers that is twenty-one platters, three RAID batteries, and six 460 watt hot-swap power supplies. Each chassis was about forty pounds and twenty-seven inches deep. HP rates the platform to 95°F. Sarasota in July does not care what HP rated.

They earned the electricity. A used enterprise server was the cheap way, a decade ago, to get ECC, a hardware RAID card, remote management, and enough RAM to be careless about how many guests you started. VMware sat on them for years. Proxmox replaced VMware last year without asking the hardware to change. The hardware changed when the guest list did.

The Replacement Is a Desktop, on Purpose

An OptiPlex 7090 Small Form Factor is a 2021 office PC. The one carrying the guests has a single Intel Core i7-10700: eight cores, sixteen threads, 2.9 GHz, a turbo of 4.8, 65 watts, Comet Lake on a 14 nanometer processor. That is one generation of Intel and about ten years past Westmere. Per thread it is in a different league. Three old boards still had more total threads — seventy-two against sixteen on one new machine, or thirty-two if both desktops are busy. Thread count was the point when the cluster was a hotel. It is not the point for a mail server, a web proxy, a reflector, and a Minecraft world, all of which spend most of their life waiting, and want one fast core when they don’t.

Memory on that node is 64 gigabytes of ordinary DDR4, four 16 gigabyte sticks, rated 2666 and running at 2400. No ECC. That is a real concession, and I will come back to it. Disk is the other direction entirely: a 500 gigabyte WD Green NVMe drive for the operating system, and a 1 terabyte WD Green NVMe for the virtual disks. Nothing seeks. Nothing spins. There is no RAID card and no cache battery. The network card is the one piece of the old design I kept on purpose, an Intel I350 with two gigabit ports, bonded, same idea as the two Broadcom ports that used to be the server uplink. The onboard desktop NIC sits unused. A hypervisor can move into a smaller box without getting cute on the network.

Physically the new machines are a different object. Dell’s small chassis is about eleven inches deep and under thirteen pounds at the heavy end of the sheet. One power cord. No out-of-band controller. No hot-swap bay. You can pick one up. You could not pick up a DL360 without planning the route out of the rack.

Three DL360 G7s OptiPlex 7090 SFF
Era Chassis 2010, CPUs 2011. In service here 12+ years. 2021 small desktop. i7-10700, Comet Lake.
CPU 2× Xeon X5675 per server. 12 cores / 24 threads. 95 W each. Turbo 3.46 GHz. 1× Core i7-10700. 8 cores / 16 threads. 65 W. Turbo 4.8 GHz.
Memory 192 GB DDR3-1333 ECC per server. 576 GB across the cluster. Twelve DIMMs each. 64 GB DDR4 on the node running the guests. Four DIMMs. Not ECC.
Disk 7× 600 GB 10K SAS. RAID 50 plus a hot spare. Smart Array P410i and a cache battery. Two NVMe drives. 500 GB boot, 1 TB for guests. No RAID card.
Power Two 460 W hot-swap supplies per server. Six supplies in the cluster. One cord. One internal supply. No management processor beside it.
Measured draw One server, light load, garage inlet 88–91°F: about 160 W settled, 205 W over the evening, spike 396 W. CPU package about 11–17 W with the remaining guests running. No wall meter in the chassis.

The Watts Were the Point

I don’t want a brochure number for this. The last DL360 that still had a working management controller reported its own power. On the evening of September 6, 2026, it logged 252 readings over five hours. The air going in the front of the chassis was between 88 and 91 degrees. The server averaged 205 watts, settled near 160 once the evening eased off, and spiked to 396. It was not busy. By then it was down to a handful of small virtual machines. A dual-processor board from 2011, twelve sticks of DDR3, and seven 10K SAS drives simply do not know how to sit still. HP’s own SPECpower run on a DL360 G7 with this same Xeon showed an active idle around 56 watts, in a lab, with a lab’s disk count and a lab’s air conditioning. A garage, a full memory load, and a shelf of spinning SAS is a different machine. The meter said so.

Three servers at that settled floor are 480 watts, every hour. That is about 4,200 kilowatt-hours a year. Florida Power & Light’s published typical residential bill for 2026 is $136.64 for 1,000 kilowatt-hours, roughly 13.7 cents. Call it about $575 a year to let three old servers idle the way one idle server behaved in September. Use the evening average instead of the floor, 205 watts times three, and it is about 5,400 kilowatt-hours and about $740. Those are the honest bounds of a measurement I actually have. July would not have been the cheap end of that range. Hotter inlet air makes the one-rack-unit fans spin harder, and fans are a load.

The OptiPlex has no meter in the power supply, so I will not invent a wattage at the wall. What I can read is the processor’s own energy counter. With the remaining guests running — mail, monitoring, the sites, Minecraft, the reflectors — the chip was drawing about 11 to 17 watts, and the memory controller’s share was about 2. Dell’s environmental sheet for this small chassis, on one of the configurations they tested, lists about 21 watts at a long idle and 26 at a short idle, and 178 watts only if you pin the processor. This box has more memory than a bare office config, plus two NVMe drives and a dual-port server NIC, so the outlet will sit above that idle line. It will not sit anywhere near 178 while this house is the workload. Even if I assume a generous 60 watts at the wall for each OptiPlex, a pair is 120 watts. Set that against 480 to 615 for the three servers and the difference is a few hundred watts, every hour of the year. At the same FPL bill that is on the order of four to six hundred dollars, and that is before counting the spikes toward 400 watts or a July afternoon.

Almost every watt a computer draws becomes heat in the room. The rack was a space heater that also answered email.

The Garage Was the Cooling System

Four hundred and eighty watts is about 1,600 BTU an hour. Six hundred and fifteen is about 2,100. A ordinary plug-in space heater is 1,500 watts. Three DL360s, at the only load I measured, were a third to two-fifths of one of those heaters, left on every hour of every day, in a garage that is not air conditioned. The monitoring write-up already calls that room a datacenter with a straight face. The temperature graph is the garage. In July the air out there reaches the mid-90s. HP’s operating range for the DL360 G7 ends at 95°F. On a mild September evening the inlet on the server was already 91. The machines were living at the top of their own spec sheet, and they were part of the reason.

A one-rack-unit server cools itself with a row of small, fast fans. Hotter air in the front means those fans speed up. Faster fans draw more watts. More watts leave as hotter air out the back, into the same room, which is the air the next server breathes. That loop is why a single fan failure on the hottest afternoon was a credible way to lose a node, and why fan health and inlet temperature were alerts I actually wanted to be woken up for. The OptiPlex has one chassis fan, sized for a 65 watt desktop processor, in a case that is not trying to be an inch and a half tall. It adds a little heat.

I don’t have a before-and-after thermometer study, and I won’t fake one. I have the inlet temperature while a DL360 was running, the watts it was dumping into that air, and the fact that twenty-one 10,000 rpm drives and six power supplies are no longer in the room. The Pi that lives out there for the radio work, and anything else that has to spend a Sarasota summer in that garage, gets a lower starting temperature. That matters more than the dollar figure. The dollar figure just makes it obvious.

Solid State, and Where the Safety Went

The disk change is the other half of the heat, and the half that changes how a failure works. A 10,000 rpm SAS drive is a small motor that never gets to stop. Several watts each, seven per server, twenty-one in the cluster, plus the RAID controller, plus a spare that existed to be ready rather than to store anything. Seek times are measured in milliseconds. An NVMe drive answers in microseconds, draws about a watt or two unless you are hammering it, and has no bearing to dry out in a 90-degree room. The guest disks sit on a thin pool on the 1 terabyte drive. Boot is the other stick. There is no array to degrade.

There is also no hot spare and no second power supply. If the data NVMe dies, nothing rebuilds. The recovery plan is the backup server, which was never one of the DL360s: a small Protectli appliance, a USB disk, and a daily copy of those backups off to Backblaze. A dead desktop drive is a restore, not a bay swap at 11 p.m. That is a deliberate trade. It is the right trade for a house whose guest list fits on one fast SSD. It would have been the wrong trade when the cluster was full of client machines and a rebuild overnight was cheaper than an outage.

The same honesty applies to memory and to lights-out management. ECC on the DL360s corrected the kind of soft errors you collect when a server runs for years. A desktop does not. I accepted that because the guests are mine, the backups exist, and 192 gigabytes of correcting memory per node had become a heater for 13 gigabytes of actual use. The iLO controllers were the other enterprise habit: a little computer inside the computer, on its own address, able to report power and fans when the operating system could not. They also drew power, needed patching, and were one more thing on the network. The new machines have nothing in that role. Monitoring reads the processor temperature, the chassis fan, the Intel NIC, and the NVMe health counters from the operating system itself. If the OS is dead, the box is dead, and I walk out to the garage. That is fine. The garage is attached to the house.

What This Is Not

This is not a claim that a small Dell is a better server than a DL360. For the job those machines were bought for — a lot of ECC, a lot of guests, a spare disk already spinning, two power supplies, a controller that answered when the host did not — the DL360 was the right shape of used hardware, and it did that job for more than twelve years. VMware and then Proxmox both ran on it without drama that belonged to the brand on the bezel. The bezel just stopped matching the work. A cluster named for a consulting practice does not need three enterprise servers once the practice’s virtual machines have gone home.

What replaced them is smaller, faster per thread, silent by comparison, and built out of solid-state storage and a power budget that looks like an office PC because it is one. The electric bill drops by hundreds of dollars a year on the measurement I have. The garage loses a heater it never asked for. The failures that remain are boring ones, and the backups were already the real safety net.

The Bottom Line

Three dual-Xeon DL360 G7s, seventy-two threads, 576 gigabytes of DDR3, and twenty-one 10K SAS drives were the right cluster for a pile of client virtual machines. They were an expensive way to heat a Sarasota garage once that pile was gone. A pair of OptiPlex 7090 desktops, one fast eight-core chip, NVMe, and a couple of gigabit ports in a bond, runs what this house actually still runs. The meters said one old server wanted 160 to 200 watts just to be awake. The new processor, doing the whole remaining job, wanted about 15. I will take that trade.