Every air-cooled desktop PC is running a quiet thermal war between its two biggest heat sources. A new fluid dynamics simulation by Inviscid AI quantifies something builders have long suspected but rarely been able to measure: the hot air your GPU exhausts into the chassis is quietly raising the temperature of the air your CPU cooler draws from — and front-panel intake fans, the conventional fix, only partially solve the problem.

The simulation, run using Inviscid AI’s open-source OpenReynolds tool, found that routing supplemental intake air through a case’s floor — directly upward toward the CPU cooler — rather than relying solely on front-panel fans reduced the CPU cooler’s intake air temperature by as much as 13°C (23°F) compared to a conventional front-intake, positive-pressure layout under simulated gaming loads. The original finding was first reported by HKEPC, a Hong Kong hardware outlet, and subsequently covered by Guru3D.

That 13°C gap is not a minor thermal variance. Under sustained loads where high-end CPUs consume 150 to 250 watts or more, a 13°C (23°F) reduction in intake air temperature translates directly into lower processor package temperatures, slower fan speeds for equivalent cooling, and meaningfully more headroom for performance under heat-limited conditions.

How Front Intake Fails: The GPU Exhaust Recirculation Problem

The conventional front-to-back, positive-pressure airflow path has earned near-universal endorsement in PC building guides for good thermodynamic reason: fresh ambient air enters from the front before touching any heat-producing component. In a typical tower build, front fans pull air through the front mesh, across the graphics card, and toward the CPU cooler and rear exhaust. Keeping pressure slightly positive — more air in than out — also discourages dusty ambient air from being drawn through unfiltered chassis gaps.

The problem the simulation surfaces is structural, not easily solved by adding more front fans. In modern ATX and mid-tower builds, the GPU sits directly below or adjacent to the CPU cooler’s own intake fan path. As the graphics card exhausts its hot air upward and outward into the chassis interior, that heated air column rises — mixing with the air the CPU cooler is simultaneously drawing from. Front intake helps, but the GPU sits between the front intake and the CPU cooler, meaning some fraction of GPU exhaust air inevitably contributes to what the CPU cooler actually draws in.

Inviscid AI’s simulation modeled this precisely. In the conventional positive-pressure front-intake configuration, HKEPC’s simulation data show CPU cooler intake air temperature climbed from an ambient 22°C (72°F) at the start of a simulated load to 35°C (95°F) — a 13°C (23°F) rise driven by GPU exhaust recirculation into the CPU cooler’s air supply. The simulation’s heat maps showed orange and red high-temperature air concentrated in the GPU rear zone and the upper chassis area — exactly the region the CPU cooler draws from.

This finding is directionally consistent with what the independent zone cooling analysis community has measured in physical builds. Builders who use thermal zone separation — giving the CPU and GPU distinct, non-overlapping intake air sources — report CPU temperature reductions of 6 to 12°C (11 to 22°F) under combined CPU and GPU stress testing compared to shared-airflow configurations.

What Bottom Intake Actually Does (and Why the Mechanism Matters More Than the Volume)

The bottom-intake configuration tested in the simulation did not simply add more total airflow volume to the system. The mechanism is more specific: fans mounted on the case floor draw cool ambient air from beneath the chassis and direct it vertically upward, delivering thermally clean air directly toward the CPU cooler’s intake path before it enters the GPU exhaust mixing zone.

In the simulation, this bypassed the recirculation problem almost entirely. HKEPC’s complete simulation data show CPU cooler intake temperature held at 22°C (72°F) — room temperature — throughout the simulated load cycle, rather than climbing to 35°C (95°F) as it did under the conventional front-intake layout. GPU bottom intake air temperature was held to 26°C (79°F) versus 30°C (86°F) in the conventional setup, an improvement in GPU thermal conditions as well.

The thermal gain comes from geometry, not fan count. Bottom-mounted fans establish a vertical air column that preempts the GPU exhaust mixing effect rather than fighting it downstream. That is why the finding is mechanistically significant, not just empirically useful: it identifies the recirculation path as the bottleneck, not the total volume of air in the system.

Does AI Do CFD Now? What Is Happening Here

The Inviscid AI study does double duty: it delivers a thermal engineering finding, and it demonstrates something larger — that an AI agent can now autonomously set up, execute, and deliver publishable CFD analysis on a consumer hardware question without a CFD-trained engineer managing the process.

OpenReynolds is a tool-use loop with seven tools — bash (for running OpenFOAM commands), write_file, read_file, job_start, job_check, job_kill, and fetch — pointed at a Linux compute instance running OpenFOAM v2512, the leading open-source CFD solver package. The OpenReynolds GitHub repository is publicly available under the MIT license. The agent writes its own case configuration files, generates and inspects mesh renders visually (it reads PNG files of the geometry before committing solver time), launches solver jobs that run as detached processes, monitors residuals in real time, and produces final figures and output files — all without the user needing to know what a blockMeshDict is or how to set boundary conditions in fvSchemes.

The architectural detail that sets OpenReynolds apart from similar AI-CFD research tools — OpenFOAMGPT from TU Ilmenau, AutoFOAM from SimuNetics, or FoamGPT from NeurIPS 2025 — is deliberate non-intervention. The OpenReynolds harness is explicitly prohibited from enforcing an order of operations, injecting checklists, blocking tool calls, or requiring approvals at any step. The OpenReynolds test suite enforces this: the build fails automatically if imperative language appears anywhere in the system prompt or briefing. The design philosophy is that a harness instructing the agent how to do CFD caps the quality of results at whatever its author knew about CFD — a deliberate inversion of the heavily structured multi-agent architectures that dominate academic AI-CFD research.

The practical consequence is that the PC case thermal study came out of the same tool that has validated its results against canonical aerospace benchmarks — the ONERA M6 transonic wing at Mach 0.84, with shock positions within 0.06 chord of 1979 wind tunnel data; a vortex-induced vibration case whose Strouhal number landed at 0.1661 against a published 0.164; a natural convection cavity with a Nusselt number within a quarter of a percent of the established benchmark. Those validation studies at tryreynolds.com do not make the PC case simulation result physically verified — but they establish the tool’s baseline credibility for the type of analysis it is being asked to perform.

Inviscid AI was founded in 2025 and is a Y Combinator Winter 2026 graduate, with over $2.4 million raised to date, according to Inviscid AI’s own studies page. The company offers a hosted version of OpenReynolds at tryreynolds.com in addition to the MIT-licensed open-source release on GitHub.

Does AI Lowering the Cost of CFD Change Hardware Journalism?

The broader implication of the OpenReynolds PC case study — which the raw simulation numbers do not state — is that engineering-grade fluid dynamics analysis is no longer gated behind commercial CFD licenses, specialist expertise, or institutional compute budgets. A builder with a well-formed question, an OpenReynolds account, and a few hours of instance time can now run the kind of thermal analysis that previously required a fluid dynamics consultant and software costing tens of thousands of dollars per seat.

That changes the evidentiary baseline for hardware claims. Thermal comparisons in case reviews have historically been limited to physical builds with temperature probes in specific, reproducible configurations. CFD simulation can model configurations that haven’t been physically built, compare multiple layout variants simultaneously, and visualize the full three-dimensional temperature field — not just the probe points a reviewer happened to instrument. The caveat is accuracy: CFD results diverge from physical reality when the model’s geometry, fan curve assumptions, or component thermal output values are inaccurate. The simulation provides a hypothesis; physical testing confirms or revises it.

For case designers and case reviewers, the Inviscid AI study offers a specific, testable hypothesis: bottom-intake fan mounts outside the PSU shroud area, delivering cool air directly upward toward the CPU cooler, should improve CPU cooler intake air temperature by a meaningful amount under sustained load. That hypothesis can be tested on real hardware. Until it is, the 13°C figure represents a simulation-backed prediction rather than a measured result.

What Builders Need to Know Before Buying or Reorganizing

Does this mean builders should immediately switch to bottom intake?

Not universally — and not without checking case compatibility first. Realizing the bottom-intake benefit requires a chassis with fan mounting positions on the floor outside the standard PSU shroud footprint, adequate clearance between the case floor and the surface it rests on to actually draw air, and bottom-facing mesh or filtration beyond what is typically reserved for PSU intake alone. Most current retail cases allocate floor vents exclusively to PSU cooling; cases with dedicated bottom fan mounts outside the shroud area are less common. Builders evaluating future case purchases now have simulation-backed data to factor into that decision.

Why hasn’t physical testing confirmed this yet?

No independently verified physical build testing corroborating the 13°C simulation result had been published as of the time of writing. That is not a disqualification — it is the next step in the research chain. CFD provides the modeling; physical testing with calibrated temperature sensors and reproducible load conditions provides the verification. The simulation is transparent: Inviscid AI’s methodology, including the full session transcript and solver files, is accessible through the OpenReynolds study library. Reviewers with the right case hardware can reproduce the test on real equipment.

What does this mean for the broader AI-in-engineering trend?

OpenReynolds is one of several AI-CFD tools that emerged in 2024–2026, alongside OpenFOAMGPT, AutoFOAM, and FoamGPT, each taking a different approach to automating engineering simulation. The practical gap OpenReynolds closes is accessibility: traditional CFD setup requires knowledge of solver selection, mesh generation, boundary condition specification, and numerical scheme choices — skills that take years to develop. An agent that handles those decisions autonomously while leaving the engineering question to the user is a meaningful change in who can commission a CFD study. Hardware enthusiasts, case manufacturers, and hardware journalists are now among the populations who can.

Frequently Asked Questions

Does bottom intake cooling actually work, or is this just simulation?

The Inviscid AI simulation shows a strong thermal benefit — up to 13°C (23°F) reduction in CPU cooler intake air temperature — with a mechanistic explanation grounded in how GPU exhaust recirculation works in tower-format PC builds. The simulation is methodologically transparent and runs on OpenFOAM, the same CFD solver used in aerospace and marine engineering validation. However, it has not yet been independently verified in physical hardware testing as of this writing. The simulation provides a strong, testable hypothesis; physical confirmation is the next step.

What cases actually support bottom-mounted intake fans outside the PSU area?

This is the key practical bottleneck. Most current mid-tower and full-tower cases allocate their floor vents and fan mounts to PSU intake, with the PSU shroud covering the majority of the case floor. Cases that specifically include additional bottom fan mounts outside the PSU shroud footprint are less common. When evaluating future case purchases, checking for dedicated bottom fan support beyond the PSU area — ideally with filtration — is the relevant specification to prioritize if bottom-intake performance matters to a builder.

How is AI being used to run these fluid dynamics simulations?

OpenReynolds connects a large language model (the AI) to a Linux compute instance running OpenFOAM, the open-source fluid dynamics solver, through seven tools: the ability to write configuration files, run shell commands, read output files (including images of mesh renders the AI uses to visually inspect its own geometry), start and monitor solver jobs, and fetch reference material from the web. The user describes the flow problem in plain language; the agent handles the technical setup from mesh generation through solver configuration to post-processing. No prior OpenFOAM expertise is required from the user. The tool is open-source under the MIT license.

What should GPU and CPU temperatures tell builders about their current airflow?

If a builder is running a high-TDP CPU alongside a high-TDP GPU and notices CPU package temperatures rising in proportion to GPU load — not just CPU load — that is a likely indicator of GPU exhaust recirculation affecting CPU cooler intake air quality. Monitoring CPU cooler intake air temperature directly, if a case has an appropriate sensor mounting point, provides more direct evidence. The threshold identified in the simulation: CPU cooler intake air can climb from ambient (22°C/72°F) to 35°C (95°F) under combined load in a conventional front-intake layout — a delta that affects every performance and noise target the cooler is tuned for.