Radiator Sizing Guide: How Much Cooling Does Your GPU Need?
Share
The One Rule That Covers 90% of Builds
Radiator sizing guides tend to overcomplicate things. There are charts with 15 variables, calculators that ask about ambient temperature and altitude, and 3-part blog series that take an hour to read. For most PC water cooling builds — including AI workstations — you need one rule:
120mm of radiator per 100W of heat, plus 120mm of overhead.
That is it. A 450W RTX 4090 needs 540mm + 120mm = 660mm minimum. Round up to the nearest common radiator size: 360mm + 360mm (720mm), or 480mm + 240mm (720mm).
This rule gives you a loop that keeps coolant temperatures under 50C at moderate fan speeds (800-1200 RPM) in a 22-25C room. If you want near-silent operation with fans at 600-800 RPM, add another 120mm. If you are in a hot climate (30C+ ambient), add another 120mm.
The rest of this article covers the details for people who want to optimize further.
Sizing for Common GPU Configurations
Single GPU Builds
| GPU | TDP | Minimum Rad | Recommended (Quiet) | Maximum (Near-Silent) |
|---|---|---|---|---|
| RTX 3090 / 3090 Ti | 350W / 450W | 360mm + 120mm | 360mm + 240mm | 480mm + 240mm |
| RTX 4090 | 450W | 360mm + 240mm | 480mm + 240mm | 480mm + 360mm |
| RTX 5090 | 575W | 480mm + 240mm | 480mm + 360mm | 480mm + 480mm |
| RTX 4070 Ti Super | 285W | 360mm | 360mm + 120mm | 360mm + 240mm |
| RTX 4060 Ti | 160W | 240mm | 360mm | 360mm + 120mm |
Note: These assume GPU-only cooling. If you are also water-cooling the CPU (adding 65-125W depending on the processor), add radiator capacity for the CPU's TDP using the same 120mm-per-100W rule.
Dual GPU Builds
| Configuration | Combined TDP | Minimum Rad | Recommended (Quiet) |
|---|---|---|---|
| Dual RTX 3090 (NVLink) | 700W | 480mm + 360mm | 480mm + 480mm |
| Dual RTX 4090 | 900W | 480mm + 480mm | 480mm + 480mm + 240mm |
| Dual RTX 5090 | 1150W | 480mm + 480mm + 360mm | External radiator recommended |
Dual GPU builds are where radiator sizing becomes a genuine constraint. 900W+ of heat in a standard mid-tower case pushes the limits of internal radiator mounting. Some builders use external radiators (MO-RA 420, or a DIY radiator outside the case connected with quick-disconnect fittings) to handle the thermal load.
For dual 3090 builds specifically, see our dual RTX 3090 NVLink cooling guide.
Radiator Thickness: Slim vs. Standard vs. Thick
Radiators come in different thicknesses, and thickness affects cooling capacity per unit of length.
| Thickness | Typical Range | Performance | Trade-off |
|---|---|---|---|
| Slim | 25-30mm | Baseline — good airflow, adequate cooling | Fits in tight cases; lower absolute capacity |
| Standard | 35-45mm | 15-25% better than slim at the same fan speed | Slightly tighter case fit; best all-around choice |
| Thick | 55-65mm | 25-40% better than slim at high fan speeds | May not fit with fans in many cases; needs high fan speed to utilize the extra depth |
The Thickness Trap
Thick radiators (55mm+) have more cooling surface area, but they require higher airflow to push air through the dense fin stack. At low fan speeds (600-900 RPM), a thick radiator may perform only marginally better than a slim one — the air simply does not penetrate deep enough to reach the inner fins.
For AI workstation builds where silence is a priority (fans at 600-1000 RPM), a 30-40mm radiator with more surface area (i.e., a longer radiator) is better than a shorter, thicker radiator. A slim 480mm outperforms a thick 360mm at low fan speeds in most configurations.
FormulaMod Radiator Options by Thickness
30mm (slim):
- Barrow Dabel-30a 240mm — 30mm slim, 14 circulation channels, fits almost any case
- Barrow Dabel-30a 360mm — same slim profile, excellent for space-constrained builds
- Barrow Dabel-30a 480mm — maximum surface area in a slim package
- Bykski RC Series 30mm — available in 120/240/360/480mm lengths
40mm (standard):
- Bykski RC 240mm 40mm — sweet spot for secondary radiators
- Bykski RC 480mm 40mm — high capacity, fits most mid-towers in top or front mount
- Bykski RC Series 40mm — available in 120/240/360/480mm lengths
60mm (thick):
- Barrow Dabel-60d 360mm — 60mm thick, maximum density. Use only if you plan to run fans at 1200+ RPM.
Fan Configuration: Push vs. Pull vs. Push-Pull
How you mount fans on the radiator affects cooling performance significantly.
| Configuration | Description | Performance | When to Use |
|---|---|---|---|
| Push | Fans blow air through the radiator | Baseline | Default choice for most builds; fans mounted on the intake side of the radiator |
| Pull | Fans pull air through the radiator | Similar to push (within 1-2C) | When you cannot mount fans on the intake side due to clearance |
| Push-Pull | Fans on both sides of the radiator | 10-20% better than push alone | Thick radiators (45mm+) where single-side fans cannot fully penetrate the fin stack |
Practical Advice
For slim (30mm) and standard (40mm) radiators, push-only is sufficient. The added cost and noise of push-pull fans is not justified — you get more cooling from buying a larger radiator than from doubling the fans on a smaller one.
For thick (55-60mm) radiators, push-pull makes a meaningful difference and is recommended. But remember: push-pull doubles the number of fans, doubles the noise sources, and doubles the fan cost. A 360mm thick radiator with push-pull needs 6 fans, not 3.
Fans for radiator duty: Bykski 120mm fans offer good static pressure (important for pushing air through radiator fins) and are compatible with all standard radiator sizes.
Case Compatibility Reference
The most common case types for AI workstation builds and their radiator capacity:
| Case Type | Typical Radiator Support | Suitable For |
|---|---|---|
| Lian Li O11 Dynamic | Top: 360mm, Side: 360mm, Bottom: 360mm | Single GPU up to RTX 5090, dual GPU with planning |
| Lian Li O11 Dynamic XL | Top: 360mm, Side: 360mm, Bottom: 360/480mm | Any configuration including dual GPU |
| Corsair 5000D/7000D | Top: 360mm, Front: 360mm, Side: 360mm | Single or dual GPU |
| Fractal Torrent | Bottom: 360mm (front: limited) | Single GPU with limitations |
| Phanteks Enthoo 719 | Top: 480mm, Front: 480mm, Bottom: 360mm | Any configuration, maximum radiator capacity |
| Mid-tower (generic) | Top: 240-360mm, Front: 240-360mm | Single GPU |
| ITX / SFF | Usually 240mm only | Low-TDP GPUs only; mid-range GPUs with compromise on noise |
Always check clearance between the radiator + fans and other components. The most common clearance conflicts:
- Top radiator vs. tall RAM modules (especially with 40mm+ thick radiators)
- Front radiator vs. GPU length
- Bottom radiator vs. PSU shroud
- Any radiator + fans vs. motherboard power connectors or IO shields
Copper vs. Aluminum Radiators
A quick but important note: never mix copper and aluminum in the same loop. All Bykski and Barrow waterblocks are nickel-plated copper. Their Dabel and RC series radiators are copper. Mixing in an aluminum radiator from another brand creates a galvanic cell that corrodes the aluminum and contaminates the entire loop.
If you see a very cheap radiator from an unknown brand, check the material. Aluminum radiators are significantly lighter than copper ones of the same size — if a 360mm radiator weighs under 400g, it is almost certainly aluminum.
All radiators linked in this guide are copper and compatible with copper/nickel waterblocks.
How to Tell If You Need More Radiator
Already have a loop running? Here is how to check if your current radiator capacity is adequate:
- Check coolant temperature. Use a Bykski temperature/flow monitor or an inline temperature sensor. Under sustained AI load in a 22-25C room, coolant should stabilize under 50C. If it exceeds 50C, you need more radiator or higher fan speed.
- Check delta-T. Delta-T is the difference between coolant temperature and ambient room temperature. Under 15C delta-T at moderate fan speed = excellent. 15-20C = good. Above 20C = you are radiator-limited.
- Listen to your fans. If fans need to run above 1200 RPM to keep coolant under 50C, adding radiator surface would allow lower fan speeds at the same thermal performance.
Radiator Order in the Loop: Does It Matter?
Short answer: barely.
The temperature difference between "GPU → radiator → pump" and "GPU → pump → radiator" is typically 1-2C. Coolant temperature equalizes quickly across the loop because the water volume moves through all components many times per minute.
What matters more: place the radiator(s) where case airflow works best. If your case has a mesh top panel, the top radiator mount will exhaust heat efficiently. If the top is solid, use front-mount intake (pull cool outside air through the radiator).
For builds with two radiators: use one as intake (pull fresh air in through the radiator) and one as exhaust (push hot air out through the radiator). This creates positive airflow through the case and keeps all components — not just the water-cooled ones — at reasonable temperatures.
Quick Decision Matrix
| Your GPU | Your Priority | Buy This |
|---|---|---|
| RTX 4060 Ti / 4070 | Any | Single 240mm or 360mm radiator |
| RTX 3090 / 4070 Ti Super | Quiet operation | 360mm + 240mm |
| RTX 4090 | Quiet operation | 480mm + 240mm (minimum 360mm + 240mm) |
| RTX 4090 | Near-silent | 480mm + 360mm |
| RTX 5090 | Quiet operation | 480mm + 360mm |
| Dual 3090 | Quiet operation | 480mm + 360mm (minimum) |
| Dual 4090 | Quiet operation | 480mm + 480mm |
Common Radiator Sizing Mistakes
Understanding what goes wrong helps more than theoretical perfection. Here are the radiator sizing errors that cause the most grief in AI workstation builds.
Sizing for the Wrong TDP Number
GPU manufacturers list TDP (Thermal Design Power) as a single number, but the actual power draw varies with workload. An RTX 4090 is listed at 450W, but under sustained AI inference it may pull 380-430W depending on the model and power limit setting. Gaming pulls 300-380W on average with spikes to 450W.
For radiator sizing, use the sustained power draw under your actual workload, not the peak TDP. If you run Ollama inference, monitor power draw with nvidia-smi for 10 minutes and use that number. You will often find you need slightly less radiator than the TDP number suggests — but size for the higher number anyway, because the extra capacity means quieter fans.
Forgetting CPU Heat Load
If you plan to add the CPU to your loop later (a common upgrade path), size your radiators for the combined heat load from day one. A typical AI workstation CPU (AMD Ryzen 9 or Intel i9) adds 125-170W of heat. If you sized your radiators for a 450W GPU only and later add a 125W CPU, you are suddenly 125W short — meaning either louder fans or higher coolant temperatures.
Ignoring Ambient Temperature
The 120mm-per-100W rule assumes a 22-25C ambient room temperature. Radiators reject heat by transferring it from coolant to air. The closer your coolant temperature is to the air temperature, the less efficient the transfer becomes. In a warm room (28-32C), you need roughly 20-30% more radiator surface to achieve the same coolant temperature. This matters for AI rigs in garages, server closets, or un-air-conditioned rooms during summer.
Buying Slim When You Have Space for Standard
Slim (30mm) radiators are great for tight cases, but if your case can accommodate a 40mm radiator in the same spot, the thicker radiator provides 15-25% more cooling capacity with the same fans. Do not choose slim just because it is cheaper — the price difference between a 30mm and 40mm Bykski radiator is $5-10, but the cooling difference is significant.
Maintenance Considerations for Radiators
Radiators require minimal maintenance compared to other loop components, but they are not zero-maintenance.
- Dust accumulation on fins: Over 6-12 months, dust builds up between the radiator fins and fan blades, reducing airflow by 10-20%. Clean radiator fins with compressed air every 6 months. Remove fans for thorough cleaning if dust is heavy.
- Internal flushing: New copper radiators may contain flux residue, solder particles, or manufacturing oils. Before first use, flush with distilled water until the water runs clear. For established loops, flushing during coolant changes (every 6 months for AI rigs) keeps internal channels clear.
- Corrosion check: During coolant changes, inspect the coolant for discoloration. Green tint indicates copper corrosion (often from mixed metals). Brown particles indicate severe corrosion. See our maintenance guide for the complete inspection procedure.
Browse all radiator options in the AI Workstation Cooling collection. For a complete build walkthrough with specific radiator recommendations, read our RTX 4090 water cooling guide or the cost breakdown by tier.
Related Articles
12VHPWR Safer RTX 5090 Build: Why Water Cooling Lets You Skip the Cable Drama
12vhpwr · 2026 · cable safety · rtx 5090 · safer build · shunt resistor · undervolt · water coolingDual RTX 3090 NVLink for 70B LLMs: The Cooling Guide
70b llm · dual rtx 3090 · llama 70b · multi gpu · nvlink · ollama · sovereign ai · water cooling10 Beginner Mistakes in Your First Custom Water Cooling Loop
beginner mistakes · coolant guide · drain port · first custom loop · fitting guide · leak test · radiator sizing · water cooling tips
Ollama Hardware Cooling Guide: Keep Your GPU Cool for Faster Inference
gpu cooling · hardware guide · inference speed · local llm · ollama · rtx 3090 · rtx 4090 · thermal throttling · water cooling