Zach Dee

Custom liquid-cooling builds & hardware diagnostics
Chestnut Mound, TN
615-359-7000
deezach.1993@gmail.com

Twelve custom builds published between 2019 and 2021 — one selected as a featured build, one sponsored by EVGA — plus the machine I run today. Three are documented here in full. Each was planned around a thermal target, assembled by hand, leak-tested before power, and validated under sustained load with published results — including the parts that failed and what it took to resolve them.

Builds
13
Featured
1
Manufacturer sponsored
1
Span
2019–26

GPU thermal rise by cooling method

Every published build carries its own idle and load temperatures, so the whole set can be read as one experiment. Plotted below is delta-T — the rise from idle to sustained load, calculated from the idle and load temperatures recorded for each build — for each graphics card across the twelve that have them. Builds are numbered chronologically throughout; the full index sits below.

Custom loop

7 builds
083900X / RTX 2080 Ti · HardlineDec 2019
12 °C
029900K / RTX 2080 Ti · Soft tubingApr 2019
14 °C
103950X / 2× RTX 2080 Ti · Hardline PETGMar 2020
15 °C
073900X / RTX 2080 Ti · Soft tubingNov 2019
17 °C
039940X / RTX 2080 Ti · Soft tubingOct 2019
18 °C
1110900K / RTX 2080 Ti · Soft tubingJul 2020
20 °C
063900X / RTX 2080 Ti · Hardline + soft runsNov 2019
21 °C

Air or AIO

5 builds
04R5 1400 / RX 580 · Stock airOct 2019
23 °C
139800X3D / RTX 5080 · All airJun 2026
24 °C
053900X / RTX 2080 · 240 mm AIOOct 2019
31 °C
092600X / GTX 1660 Super · Tower airFeb 2020
33 °C
1210900K / RTX 3090 · Factory GPU airJan 2021
54 °C
0 °C3060 °C rise

The two groups do not overlap. Every card on a custom loop landed between 12 and 21 °C of rise; every card on air or an all-in-one landed between 23 and 54 °C. Same builder, same instrumentation, twelve machines over two years — which is what makes the comparison worth anything. Within the loop group, tubing type tracks nothing — radiator area and block contact drive the spread, which is why I moved to soft tubing for maintainability without giving anything up thermally.

What this comparison does not show. The air group's four cards range from an RX 580 to an RTX 3090 — very different heat outputs, so the spread within that group says more about the cards than the coolers. The comparison that holds is between groups, and the cleanest single pair is builds 05 and 07: same CPU, five weeks apart, AIO against a loop, 31 °C versus 17 °C. Build 13's figures come from a lighter session than the rest and are not directly comparable to the 2019–21 numbers.

Build records

10

3950X / Dual RTX 2080 Ti

featured

Built 10 Mar 2020 · published 12 Mar 2020 · $6,494 · hardline PETG

A dual-GPU workstation on a full hardline loop. Both graphics cards were stripped to bare boards and converted to waterblocks with matched backplates, then linked over NVLink. Three radiators carried the combined CPU and dual-GPU load in a case whose radiator clearance was tight enough that the 480 mm had to be relocated to the rear chamber and the run planned around the wiring that crossed it.

Compute
AMD Ryzen 9 3950X (16C/32T) at 4.2 GHz · ASRock X570 Aqua · 32 GB DDR4-3600 CL17
Graphics
2 × Asus ROG Strix RTX 2080 Ti over NVLink, both converted to EK Quantum Vector waterblocks with nickel backplates
Loop
Hardline rigid tubing, 16 mm OD · EK Torque compression, 45° and 90° rotary fittings · EK Quantum Kinetic FLT 360 D5 pump/reservoir
Waterblocks
2 × EK Quantum Vector Strix RTX 2080 Ti (nickel/plexi) with nickel backplates · CPU, VRM and chipset on the X570 Aqua integrated monoblock
Heat rejection
EK CoolStream XE 360 (thick) + SE 360 + SE 480 · Noctua F12 industrialPPC-3000 and chromax on all three
Power
EVGA SuperNOVA 1300 G2 · Phanteks Enthoo 719
CPU idle
31 °C
CPU load
67 °C
GPU load
42 °C
GPU clock
2.205 GHz

Fault log

  • Coolant separation. The premixed opaque coolant broke and separated in the loop. Flushed, diagnosed as a batch problem rather than a contamination issue, and replaced with a different coolant from the same line that had run clean across half a dozen prior builds. resolved in-house
  • Motherboard block damage. The monoblock's fin stack abraded the acrylic covering over the CPU and chipset sections. Attempted a controlled teardown to inspect, found the block assembly non-serviceable, and escalated to the manufacturer. ASRock RMA
  • GPU block lighting failure. Addressable RGB on one waterblock failed immediately after assembly, having worked in the previous system. Isolated to the block rather than the controller or header by swapping both, then escalated. EKWB RMA
11

10900K / RTX 2080 Ti

Built 27 Jul 2020 · published 30 Jul 2020 · $3,288 · soft tubing

The deliberate move off hardline. After several rigid-tube builds, the maintenance cost had become the limiting factor: every hardware change meant draining and re-bending a full set of runs, and a daily-driver machine can't be down that long. This build used matte black soft tubing instead and reached a higher CPU clock at a lower load temperature than the build before it. The recommendation I published with it — soft tubing unless you keep a backup system — went against the aesthetic preference I'd been building on for years.

Compute
Intel Core i9-10900K at 5.2 GHz · Asus ROG Maximus XII Hero · 32 GB DDR4-3600 CL17
Graphics
Asus ROG Strix RTX 2080 Ti on an EK Quantum Vector block with nickel backplate
Loop
EK-Tube ZMT matte black soft tubing, 10/16 mm · EK Velocity acetal/nickel CPU block · EK Torque nickel fittings
Waterblocks
EK Velocity acetal/nickel CPU block · EK Quantum Vector Strix GPU block with nickel backplate
Circulation
EK Quantum FLT 360 (DDC) plus a D5 tube reservoir — the DDC pump failed in service and was replaced under ticket
Power
EVGA SuperNOVA 1300 G2 · Phanteks Enthoo 719
CPU idle
28 °C
CPU load
60 °C
GPU load
47 °C
CPU clock
5.2 GHz

Fault log

  • Memory qualification. The board rejected a 64 GB CL14 kit that ran fine elsewhere. Rather than fight it, I sourced a kit from the board's qualified vendor list, which booted on XMP at 1.35 V without adjustment. Later tested three B-die kits against each other and published a correction when the two-stick result contradicted my first recommendation. resolved in-house
  • Pump failure. The DDC unit in the pump/reservoir combo failed. Opened a support ticket for the D5 replacement ring so the existing top could be reused rather than replacing the whole assembly. EKWB ticket
12

10900K / RTX 3090

Built 23 Dec 2020 · published 7 Jan 2021 · $4,377 · soft tubing

Rebuilt after a teardown accident destroyed three components at once. Documented publicly rather than quietly replaced, including the root cause, because the failure mode — a block assembly still tethered by zip-tied cabling when it came free — is the kind that repeats. Ran on mixed cooling by necessity: CPU on water, the new RTX 3090 on its factory air cooler while a waterblock was unavailable, which is what produces the 54 °C outlier in the chart above.

Compute
Intel Core i9-10900K at 5.2 GHz · Asus ROG Maximus XII Hero · 16 GB TEAMGROUP T-Force Xtreem DDR4-4133 CL18
Graphics
EVGA RTX 3090 FTW3 Ultra (24 GB) on factory air, pending block availability
Loop
Watercool Heatkiller IV Pro acetal CPU block · Heatkiller Tube 200 reservoir · Black Ice Nemesis 360GTS radiator · ZMT soft tubing with EK Quantum Torque STC-10/16
Waterblocks
Watercool Heatkiller IV Pro acetal CPU block · GPU uncooled by the loop, on its factory air cooler
Storage
2 × Corsair MP600 2 TB Gen4 NVMe · 2 × Kingston UV500B 960 GB
Power
EVGA SuperNOVA 1600 P2 Platinum · Phanteks Eclipse P500A
CPU idle
26 °C
CPU load
56 °C
GPU load
80 °C
GPU cooling
air

Fault log

  • Teardown accident. Two linked GPU blocks fell during disassembly while still connected and zip-tied to the board, tearing components off the motherboard side. The board would not POST afterward. Documented the sequence publicly so others could avoid the same mistake. root cause published
  • Swollen chokes, both graphics cards. Found identical swelling and residue on the same component in the same position on both RTX 2080 Tis, with a burnt-electronics odour and no coolant leak — ruling out the loop as the cause. Both cards were submitted to the manufacturer for repair. Asus repair
  • Thermal ceiling on air. The 3090 ran 75–83 °C in normal gaming sessions. Tested industrial-pressure fans in multiple orientations against the card, measured negligible improvement, and concluded the limit was the cooler rather than case airflow. diagnosed, block sourced
13

9800X3D / RTX 5080

current

Built 21 Jun 2026 · in daily use · all air

The machine I run now, and the one build in the set with no loop in it at all. After twelve builds of increasingly involved liquid cooling, a high-airflow mesh case, a twin-tower air cooler and a three-fan graphics card hold a 16-core-class workload in the low fifties. Nothing here needs water, and putting water in it would buy nothing but maintenance — which is the conclusion the previous twelve builds were already pointing at.

Compute
AMD Ryzen 7 9800X3D (8C/16T) · Gigabyte X870E AORUS ELITE WIFI7 rev 1.2 · 32 GB DDR5-6000 CL30-36-36-76 at 1.35 V
Graphics
Gigabyte RTX 5080 WINDFORCE SFF 16 GB GDDR7 · core offset +400 MHz under a 0.900 V cap, which measured faster and 17 W cheaper than the same clock at 0.950 V
Cooling
be quiet! Dark Rock Elite twin-tower air cooler, 2 × Silent Wings 135 mm · Thermal Grizzly Duronaut, chosen over Kryonaut because Kryonaut pumps out on high heat-density dies
Airflow
Fractal Design Meshify 3, 3 × 140 mm front intake and 1 × 120 mm rear exhaust on a Fractal Adjust Pro Hub with per-port PWM translation curves, driven from a single board header
Storage & power
Gigabyte AORUS Gen4 7300 2 TB NVMe in the CPU-attached M.2 slot · be quiet! Straight Power 12 1500 W Platinum
CPU idle
43 °C
CPU load
52 °C
GPU load
51 °C
GPU peak
250 W

Fault log

  • Fans ignoring every PWM command. Case fans stuck at full speed regardless of curve. Traced past the software — two controllers were writing the same SuperIO chip and fighting, producing oscillations that mimic a hardware fault — and then to a loose 4-pin on top of the front panel. The hub was healthy on USB the whole time, because its USB side and its PWM input are separate paths. resolved in-house
  • Repeated hard resets. Isolated to mains power rather than the machine, and resolved with a UPS rather than by replacing parts. root cause external
  • Ring-0 drivers left by an uninstaller. Removed a vendor utility and found it had orphaned two kernel-level helper drivers, one of them left behind by its own uninstaller. Verified the removal against the registry and confirmed Memory Integrity and the vulnerable-driver blocklist had been enforced throughout. verified clean

Full build index

All thirteen machines in chronological order. Numbers 10 to 13 are documented in full above. Two of the twelve were built for other people to a set budget rather than for myself, and one was sponsored by EVGA, who supplied the graphics card, power supply, case and motherboard. Delta-T columns are the idle-to-load rise each build recorded at the time.

#PublishedCPUGPUCooling / waterblocks CPU idle → loadGPU idle → loadNotes
01Mar 2019Core i7-8700K @ 5.0 GHzRTX 2080 Strix280 mm AIO · Corsair H115i——No temperatures recorded
02Apr 2019Core i9-9900K @ 5.2 GHzRTX 2080 Ti StrixSoft tubing · EK Velocity + Vector Strix24 → 45Δ2124 → 38Δ14CPU lapped; best result in set
03Oct 2019Core i9-9940X @ 4.6 GHzRTX 2080 Ti FTW3Soft tubing · Heatkiller IV Pro + EK Vector34 → 72Δ3830 → 48Δ18Sponsored by EVGA
04Oct 2019Ryzen 5 1400 @ 3.5 GHzRX 580Stock air38 → 58Δ2042 → 65Δ23Built and sold to a buyer's budget
05Oct 2019Ryzen 9 3900XRTX 2080 XC240 mm AIO · Fractal Celsius S2437 → 68Δ3132 → 63Δ31—
06Nov 2019Ryzen 9 3900XRTX 2080 Ti StrixHardline + soft · EK Vector Strix27 → 65Δ3822 → 43Δ21Hardline acrylic
07Nov 2019Ryzen 9 3900X @ 4.4 GHzRTX 2080 Ti StrixSoft tubing · Phanteks C350I + Glacier30 → 60Δ3026 → 43Δ17Pump air lock diagnosed pre-failure
08Dec 2019Ryzen 9 3900XRTX 2080 Ti StrixHardline · X570 Aqua monoblock28 → 55Δ2728 → 40Δ12X570 Aqua monoblock
09Feb 2020Ryzen 5 2600X @ 4.3 GHzGTX 1660 SuperTower air · Hyper 21233 → 58Δ2532 → 65Δ33Built to resell
10Mar 2020Ryzen 9 3950X @ 4.2 GHz2 × RTX 2080 TiHardline PETG · 2 × EK Quantum Vector31 → 67Δ3627 → 42Δ15Featured build; dual GPU over NVLink
11Jul 2020Core i9-10900K @ 5.2 GHzRTX 2080 Ti StrixSoft tubing · EK Velocity + Quantum Vector28 → 60Δ3227 → 47Δ20—
12Jan 2021Core i9-10900K @ 5.2 GHzRTX 3090 FTW3Soft tubing · Heatkiller IV Pro; GPU on air26 → 56Δ3026 → 80Δ54GPU on factory air
13Jun 2026Ryzen 7 9800X3DRTX 5080 WINDFORCEAll air · Dark Rock Elite43 → 52Δ927 → 51Δ24Current machine; not published to PCPartPicker

What the work involved

Diagnosis under uncertainty
Isolating a fault to the failing part before replacing anything — swapping controllers and headers to prove a lighting failure was in the block, ruling out the loop as the cause of two identical component failures, testing fan configurations before concluding a cooler was the limit rather than the case, and catching a pump running dry on trapped air during a radiator swap before it failed.
Measurement over assumption
Every build published with idle and load temperatures, clocks, and memory speeds. Overclocks tuned toward a voltage and thermal ceiling rather than a headline number, after confirming the real-world difference was negligible. The earliest build in the set records a radiator moved to front intake that lowered CPU temperatures and raised GPU temperatures — both halves of the tradeoff, not just the half that flattered the change.
Working across vendors
Blocks, pumps, radiators and fittings sourced across EK, Alphacool, Watercool/Heatkiller and Phanteks, and mixed within single loops — a Heatkiller CPU block against an EK graphics block in one build, Phanteks blocks on an Alphacool pump and reservoir in another. Card-specific blocks have to be matched to the exact board revision, so this is component compatibility work, not brand preference.
Vendor escalation
Warranty cases opened and carried to resolution with three manufacturers — including diagnosing far enough first to make a specific request rather than returning a whole assembly.
Documentation
Written build logs covering configuration, results, and failures. Faults recorded with the same weight as successes, including my own errors, so the record is usable by someone else.
Building to a brief
Two of the twelve were built for other people rather than myself — specified against their budget, not mine, benchmarked against my own machines before handover, and deliberately left with an upgrade path so the next component could go in without a rebuild.
Recognised by manufacturers
One build sponsored by EVGA after a direct approach, with the graphics card, power supply, case and motherboard supplied. One build selected as a featured build by PCPartPicker.
Helping other people
Answering build and troubleshooting questions from readers — platform tradeoffs for someone else's budget rather than a restatement of my own, parts lists sent to a first-time builder worried about under-ordering fittings, and a published correction when further testing contradicted advice I'd already given.