The TOPS Number Doesn't Survive Contact With Reality

Every new AI PC ships with a headline spec: an NPU rated at 40, 50, sometimes past 100 TOPS. The pitch is simple — local LLMs, image generation, and real-time vision pipelines running smoothly on your desk, no cloud required.

Here's the catch: that number comes from a benchmark run in near-laboratory conditions. The chip executes a short synthetic burst, alone, on a cold rig, with the CPU idling and the GPU powered off entirely. No background load, no competing heat sources, no real-world mess.

That's not how anyone actually uses a workstation. Load a local Llama model, kick off a multi-step diffusion pipeline, or run a vision model continuously for ten minutes, and a completely different picture emerges — the GPU is pushing parallel matrix operations, VRAM is streaming gigabytes of model weights per second, the CPU is juggling background processes and IDEs, and the NPU is handling its own inference stream on top of all of it. Every one of those components shares the same motherboard and the same enclosed case. Heat from the GPU and CPU doesn't stay contained — it spreads straight across the board substrate into the NPU's own thermal zone.

That shared, cumulative heat load — not the number printed on the box — is what actually decides how fast your machine runs ten minutes into a session.

Throttling Explained: DVFS, Not a Defect

The first time a workload grinds to a crawl mid-session, it's easy to assume something's broken. It isn't. What's happening is Dynamic Voltage and Frequency Scaling (DVFS) — a built-in safety mechanism, not a malfunction.

Every chip has a hard thermal ceiling: roughly 95°C for AMD silicon, 100°C for Intel chips. Once onboard sensors detect the die is closing in on that limit, firmware steps in immediately — clock speed drops, voltage drops, and the chip trades speed for staying intact.

A single throttling event costs nothing. The real concern is cumulative: a workstation that spends month after month in a near-constant throttled state puts real mechanical stress on the system. Repeatedly cycling between a hot die and room temperature expands and contracts solder joints and thermal interface material — over roughly 12 to 18 months of sustained heavy use, that cycling can dry out thermal paste into a brittle crust or stress solder joints under BGA packages. None of that happens from one hot session. It happens from making "hot" the default state.

Why AI Workloads Punish Hardware Harder Than Games Do

Gaming and AI inference stress the same silicon in completely different ways, even at similar wattage.

Games are naturally bursty — an intense scene spikes GPU load, a loading screen or menu gives it a moment to recover. That rhythm gives coolers small windows to shed heat before the next spike.

AI inference doesn't offer that break. A local LLM generating tokens, or a batch image job running end-to-end, keeps tensor cores and matrix units pegged near full utilisation continuously from start to finish. In a typical 10-minute sustained inference session, the pattern looks something like this:

The drop isn't gradual. It's a cliff — latency roughly quintuples once the ceiling is crossed, turning a fast, responsive session into a stutter. In severe cases, driver timeouts (TDR) trigger and the inference process crashes outright.

Software Fixes That Cost Nothing

Before opening the case or spending money, these two changes solve a surprising amount of the problem.

Undervolt first. Undervolting keeps your stock clock speed but supplies slightly less voltage than factory default. Since heat scales roughly with the square of voltage, small reductions produce outsized temperature drops. On GPUs, tools like MSI Afterburner let you flatten the voltage curve — capping around 0.900V–0.925V instead of the 1.05V+ many cards push at boost. On CPUs and integrated NPUs, Intel XTU or AMD Ryzen Master let you apply a conservative negative voltage offset, typically starting around -20mV and adjusting from there. Test every step incrementally with a real sustained workload — not a quick benchmark — before calling it stable.

Cap the power target if that feels safer. Most hardware ships tuned a few per cent past its efficiency sweet spot just to win a slightly higher benchmark score. Dialling the power limit back to around 85–90% typically costs only a small amount of peak speed while dropping sustained temperatures by roughly 8–12°C — often enough on its own to keep a long job from ever hitting the thermal ceiling.

Physical Cooling Upgrades

Once software tuning hits its limit, moving heat out of the case faster is the next lever to pull.

Fix airflow direction, not just volume. Aim for slightly positive case pressure — intake fans pulling in a touch more air than exhaust fans remove — to keep a steady stream of cool air crossing the motherboard and NPU heatsink while keeping dust out. In multi-GPU builds, leave at least two to three slots of open space between cards; mounting them flush starves the top card of air.

Replace factory thermal paste. Stock paste dries out fast under repeated 80°C+ cycles. A higher-quality compound, or a phase-change pad like Honeywell's PTM7950, stays effective far longer — phase-change materials soften under heat and re-settle into microscopic gaps on the die rather than drying into a crust.

Move to liquid cooling for multi-GPU or 24/7 setups. A single GPU with a solid air cooler is usually enough. Once you're running multiple cards or continuous fine-tuning pipelines, a 360mm+ AIO stops being optional.

Set a static fan curve. Default "reactive" curves stay quiet until 80°C, then jump straight to full speed. For sustained AI work, a flat curve — fans locked around 70–75% for the whole session — keeps airflow steady before the heatsink's thermal mass saturates.

Desktop vs. Laptop: Different Thermal Budgets

Desktops have real room to work with — undervolting, better paste, and airflow tuning can eliminate throttling almost entirely. Thin laptops are fighting physics from the start: cramming a CPU, NPU, RAM, and battery into a chassis under 20mm thick leaves heat nowhere to go. On that kind of hardware, chasing the highest TOPS number matters less than picking a laptop with genuinely good cooling — a lower-TOPS chip that holds its clocks will usually beat a higher-TOPS chip that throttles inside five minutes. If sustained workloads on a thin laptop are unavoidable, an active cooling pad and a capped power profile help more than most people expect.

FAQs

Does throttling damage my hardware? Not from a single event — that's the chip protecting itself. The risk is cumulative: months of constant heavy throttling stresses solder joints and dries out thermal paste faster than normal use.

Will undervolting void my warranty? Generally no, since it's a software change that reduces voltage and heat. Push it too far, and you'll get an instability crash, not damage — but check your manufacturer's specific policy.

Does a higher TOPS number mean better real-world performance? Not reliably. TOPS reflects a brief, cooled, isolated burst. Sustained real-world speed depends far more on how well the system handles heat than on the number printed on the spec sheet.

Is liquid cooling necessary for AI work specifically? Only past a certain point. A single GPU with decent air cooling and good case airflow is usually enough; liquid cooling earns its place with multi-GPU builds or genuinely continuous workloads.

Why does my laptop run warm even without heavy AI tasks? Background OS features — indexing, live captions, contextual suggestions — wake the NPU for small bursts constantly, which is enough to keep a thin chassis warmer than expected even at idle.

Bottom Line

Thermal throttling isn't a hardware defect — it's physics showing up exactly where marketing benchmarks leave it out. As local AI moves from occasional experiments into daily workflow, the TOPS number on the box matters less than how well your specific machine handles sustained heat. A modest undervolt, a realistic power cap, clean airflow, and a steady fan curve usually close most of that gap on their own.