Request a Quote
ACEGEEK
Looking for a PC Hardware Supplier?
Explore ACEGEEK PC cases, cooling solutions, and more. Contact our team for wholesale and bulk order inquiries.
Reviews
2026/11/18

Thermal Testing for a High-TDP CPU and GPU Running Together

Thermal Testing for a High-TDP CPU and GPU Running Together

A high-TDP CPU and GPU cooled in the same chassis share one air supply, so each component's cooling affects the other. A thermal test that runs them one at a time can give a passing result for each part while the assembled system throttles under real use. Testing them together shows how much heat the CPU cooler passes to the graphics card, and how much GPU exhaust the CPU cooler takes back in. It also shows how both react once the case has fully warmed up. This guide sets out a repeatable protocol for system integrators, brand product managers and sourcing teams. It covers power targets, load selection, ambient control, test duration, sensor logging and how to judge the results.

Why Does Testing the CPU and GPU Separately Miss Thermal Problems?

Separate tests miss the interaction between the two cooling systems. A CPU-only test shows how well the CPU cooler works when the GPU is idle. A GPU-only test shows the reverse. Neither shows what happens when a radiator, a tower cooler and a graphics card compete for the same intake air. ACEGEEK's guide to how CPU cooling and GPU airflow compete inside the same chassis explains why a layout that favors one sensor often moves the problem to the other.

GamersNexus' Cooler Master H500P radiator placement guide is a clear example. It ran the CPU and GPU under power-virus loads at the same time and reported results as temperature above ambient:

Configuration (240 mm AIO, H500P case)

CPU above ambient

GPU above ambient

What it shows

Front radiator, push-pull with front 200 mm fans

45.8°C

57.0°C

Best CPU result, but the GPU runs hotter

Front radiator, 200 mm fans moved to the top as exhaust

47°C

59°C

Top exhaust pulled air away from the GPU

Top radiator exhaust, front 200 mm fans as intake

51.8°C

50.5°C

CPU is warmer, but the two components are better balanced

Three limits apply. These are one case, one 2017 test bench and one lab's method. The configurations also changed fan roles and positions, so airflow direction was not the only variable. The temperatures are far lower than a modern flagship pairing would produce. The useful point is that the best CPU result and the best system result were different configurations. GamersNexus also stated that radiator placement should be tested case by case, not treated as a universal rule.

What Power Target Should You Test At?

Test at the power the product will actually draw, and record it. Do not test at the printed TDP. Intel's TDP documentation defines TDP as the power consumption under the maximum theoretical load. It says the figure is a design target for thermal solutions. It also says processors can exceed it for a limited time during turbo or certain workloads, such as AVX. For 12th Gen and later Intel Core processors, Intel calls this value Processor Base Power. ACEGEEK's explainer on what TDP means for CPUs and coolers covers the basic distinction between a CPU's heat output and a cooler's rated capacity.

Published ratings can sit far from sustained behavior. ACEGEEK's article on how much cooling today's high-TDP CPUs need cites an Intel Core i9-14900K at 125 W base and 253 W maximum turbo power. It cites an AMD Ryzen 9 9950X at a 170 W default TDP. Motherboard defaults decide where a given build lands between those figures, so a test that ignores power limits is not repeatable. For each run:

  • Record the BIOS power-limit settings and leave them unchanged between configurations.

  • Log CPU package power and GPU board power throughout.

  • Test at the limits the shipped product will use, not at a bench-tuned setting.

  • If you want a worst-case margin check, run a second pass at the highest limits the board allows.

Which Workloads Should Run Together?

Use two layers. A synthetic worst-case load ranks configurations and measures safety margin. A realistic mixed load decides whether the product is acceptable. A hotter test is not automatically a better test, because maximum-power synthetic loads can produce heat output that no customer workload sustains. Ranking two cases under a power virus is valid. Failing a build because it reaches a thermal limit under that load is not, since Intel's throttling guidance describes throttling as a protective mechanism. A better failure signal is throttling at modest power, or a result well above a known-good reference at the same power.

Load profile

Purpose

What it reveals

Limitation

CPU-only sustained

Baseline for the CPU cooler

Cooler capacity, mounting quality, coolant plateau

Hides GPU heat interaction

GPU-only sustained

Baseline for the graphics card

Intake path quality, GPU hotspot behavior

Hides radiator heat interaction

Combined synthetic (CPU and GPU stress tools together)

Worst-case ranking of configurations

Interaction effects, margin to throttling

May exceed real-world power draw

Combined realistic (for example, a sustained render or compute job with a 3D benchmark)

Acceptance decision

Behavior under plausible customer use

Harder to automate and repeat

The CPU-only and GPU-only runs are not redundant. They give you the baselines you need to quantify interaction, as described in the results section below.

How Do You Set Up a Repeatable Combined Thermal Test?

Step 1: Fix the hardware and the cooling settings

Use the same CPU, GPU, motherboard, memory, cooler, case configuration and BIOS settings across the comparison. Fix fan and pump speeds, and record actual RPM, because percentage settings do not guarantee identical speeds. GamersNexus manually sets case fans to fixed speeds because automatic curves reduce repeatability. Fixed-speed testing answers a case-airflow question. Testing with the stock fan curves answers a different question about the customer's experience, so run both and label them.

Step 2: Control and log ambient temperature

Room temperature changes between test days, and a result from a cool morning looks better than the same unit on a warm afternoon. Report temperature above ambient, as ACEGEEK explains in its article on why ambient-normalized temperatures matter in PC case testing. Place the probe near the case intake, away from exhaust air, and log it throughout the run. Ambient normalization does not correct for differences in CPU power, fan speed or workload, so keep those fixed or recorded as well.

Step 3: Wait for thermal equilibrium

Short runs flatter liquid-cooled systems, because the coolant and radiator warm slowly. GamersNexus' published method collected a 10-minute idle baseline. It then ran load for 1,400 seconds, which it judged sufficient for equilibrium on its own bench. Treat that figure as one lab's result, not a standard. OCCT's testing guides recommend at least 30 minutes of GPU testing to detect throttling under sustained load. They list one hour as the minimum CPU test duration for overclock validation. For your own product, plot temperature against time and set the run length from the point where the curves flatten. One illustrative criterion, which is a starting point and not a standard, is no more than about 1°C change in CPU package and coolant temperature across a 10-minute window. ACEGEEK's AIO stress-test protocol before shipping a PC covers air purging, plateau criteria and sampling tiers for liquid-cooled units.

Step 4: Log the right sensors

Sensor

Why it matters

Ambient temperature

Basis for delta T and for comparing test days

CPU package temperature and power

Separates cooling changes from power changes

Effective CPU clock and throttle flags

Shows whether a temperature actually costs performance

GPU core temperature, hotspot, memory junction

A core reading alone can hide the hottest region; OCCT's guidance lists hotspot and memory junction among the sensors to enable

GPU board power and fan RPM

Confirms comparable load; rising fan RPM signals a warmer intake

Coolant temperature and pump RPM (AIO)

Shows loop saturation and pump faults

Case and radiator fan RPM

Confirms the cooling effort was the same

Motherboard VRM temperature, rail voltages

Combined load stresses power delivery as well as cooling

How Do You Read the Results?

Compare complete configurations, and calculate an interaction penalty for each component. This is a practical metric, not an industry standard. For each component, subtract its delta T in the solo test from its delta T in the combined test. A CPU that rises 8°C when the GPU is loaded points to GPU exhaust heating the CPU cooler's intake air. A GPU that rises sharply when the CPU is loaded points to a radiator or tower cooler pushing heat into the GPU's air path. Interaction penalties show where the airflow layout is the problem. Absolute temperatures do not.

Other checks that make the comparison useful:

  • Both components together. GamersNexus argued that the GPU should be kept as cool as possible because its boost behavior responds to temperature more than a CPU's performance responds to a few degrees. That is one lab's judgment, and it depends on the chips involved, so verify it with the effective-clock data you log.

  • Radiator direction. If the CPU improves while the GPU worsens, review the trade-offs in ACEGEEK's radiator intake versus exhaust comparison, then retest with one variable changed.

  • Pass/fail limits. Define them as an allowed deviation from a known-good reference build with the same CPU, power limits, case, radiator position and fan profile. Do not use a universal temperature. Sensor uncertainty and run-to-run variation mean a 1°C gap should not decide a winner without repeat runs.

  • Power delivery. OCCT lists voltage drops under load among the signs of PSU weakness, so a crash or reboot under combined load may be an electrical fault and not a cooling fault.

What Mistakes Invalidate a Combined Thermal Test?

  • Changing several variables at once. Moving the radiator, changing fan curves and swapping the GPU in one pass makes the result unreadable. OCCT's best-practice list also recommends testing one component change at a time.

  • Testing with the side panel off. Intake restriction behind a glass or filtered panel is part of what the customer receives.

  • Using automatic fan curves in a ranking test. The system may cool better only because fans ramped faster, at higher noise.

  • Stopping before equilibrium. A short run can show a healthy CPU while the coolant is still warming.

  • Testing a different configuration from the one that ships. Radiator position, tube orientation, filters and pre-installed fans all change the result.

What Should Buyers Specify Before Requesting a Quote or Approving a Test Protocol?

A thermal test is only useful to a procurement team if it is written down and repeatable. Whether you assemble systems yourself or buy from an OEM or ODM partner, define these items before you approve a build or request a quote:

  1. CPU, GPU and power limits. Include the BIOS settings and the board power limit of the GPU.

  2. Case, cooler and radiator position. Specify the shipped configuration, including fan count, fan models and intake and exhaust roles.

  3. Load tools and profiles. Name the specific applications, the data set or mode, and whether the run is CPU-only, GPU-only, combined synthetic or combined realistic.

  4. Ambient range and probe position. State the acceptable room temperature range and where it is measured.

  5. Duration and equilibrium rule. Give minimum and maximum times and the plateau criterion.

  6. Logged data. List the sensors and require logs linked to serial numbers for warranty and RMA analysis.

  7. Reference build and allowed deviation. Agree what counts as a pass.

  8. Sampling plan. Decide which units get a long run and which get a per-unit screen.

When you request quotes, ask suppliers to put these items in writing alongside MOQ, lead time and warranty terms. A supplier that cannot describe its pass/fail criteria cannot show that its units were tested to a repeatable standard. ACEGEEK's site invites wholesale and bulk inquiries for PC cases and cooling solutions, so buyers who want to run this protocol against a specific chassis and cooling combination can start there.

Frequently Asked Questions

Should I stress-test the CPU and GPU at the same time?

Yes, in addition to separate baselines. Combined loads reveal shared-airflow effects that single-component tests hide, such as radiator heat reaching the GPU or GPU exhaust warming the CPU cooler. Run CPU-only and GPU-only tests first so you can calculate the interaction penalty.

How long should a combined CPU and GPU thermal test run?

Run until temperatures reach equilibrium, and treat published figures as starting points. OCCT recommends at least 30 minutes of GPU testing and one hour for CPU overclock validation. GamersNexus used 1,400 seconds on its own bench. Liquid-cooled systems usually need longer, so plot your own data.

Is TDP the right number to test at?

No. Intel defines TDP as power under the maximum theoretical load, and processors can exceed it temporarily during turbo. Motherboard defaults also change sustained power. Test at the power limits the product ships with, and log package power and GPU board power in every run.

Why use delta T over ambient instead of raw temperatures?

Delta T subtracts room temperature, so tests on different days or in different rooms can be compared more fairly. It does not correct for differences in component power, fan speed or workload, so those must still be fixed or recorded alongside the readings.

Does a CPU or GPU hitting its thermal limit mean the cooling failed?

Not automatically. Intel describes throttling as a protective response to temperature above the limit. Under power-virus loads, reaching that limit may be expected. Stronger failure signals are throttling at modest power, or temperatures well above a known-good reference build at the same power.

What GPU temperatures should I log besides core temperature?

Log GPU hotspot and memory junction temperatures alongside core temperature, plus board power and fan RPM. OCCT's GPU guidance lists hotspot and memory junction among the sensors to enable, because a core reading alone can hide the hottest region of the card.

Related posts