Last time, you worked through the full spec comparison table and landed on one number: at full load, each wire on a 12VHPWR / 12V-2x6 connector uses 90% of its rating, leaving 10% margin. You probably figured that as long as your card never pushes current past its rated value, that 10% holds and you're fine.
Except if you open Intel's ATX12VO design guide, there's something more startling in there: the spec says, in plain writing, that your GPU is legally allowed to spike to three times its rated power for a brief moment. Not overclocking, not a fault, not some vendor cutting corners — it's a behavior the spec itself permits, with its own name: power excursion.
This post breaks down that rule: how high your card is allowed to spike, for how long, what the power supply on the other end is required to survive, and what the ATX 3.1 revision in 2023 actually changed — and what it didn't. By the end you'll have a distinction that matters a lot: telling a spec-permitted instantaneous spike apart from an overload that shouldn't be sustained at all. Those two sound alike. They aren't the same thing, and mixing them up is exactly the misunderstanding this post untangles at the end.
The spec says it outright: your card can legally hit 3x its rating for 100 microseconds
The rule itself reads, in one sentence: "Add-in Card must at all times and concurrently adhere to power excursion limits for all time interval lengths." In plain English: at any given instant, your card has to satisfy every time window's limit at once, not just whichever one you happen to be thinking about. The catch is that different window lengths come with completely different limits.
The most extreme entry sits in Table 3-1 of the ATX12VO 2.01 spec: for any window of 100 microseconds or less, your GPU is legally allowed to hit three times its rated power.
Three times is an abstract multiplier until you put your own card's numbers into it. Say your GPU is rated for 600 W. Within 100 microseconds, it can legally spike to 1800 W. Divide by 12 volts and that's 150 amps for an instant — mostly still flowing through that 12-pin cable, split across six wires, roughly 25 amps each. What's the per-pin rating you just committed to memory in the last post? 9.2 amps. Do the division and the legal instantaneous spike comes out to about 2.7 times that per-wire rating.
And that's the best-case version, assuming the six wires split it evenly. The last post in this series covered how your GPU likely can't even measure what each of the six wires is carrying individually, and current doesn't obediently divide itself evenly on its own. If that instantaneous spike also happens to land unevenly, the wire carrying the concentration would see something higher still. Nobody has put a high-speed oscilloscope on this specific question, so this post can't give you an exact number — but there's nothing stopping it from happening.
But "100 microseconds" is only the most extreme point on this curve. Stretch the window out, and how does the allowed multiplier change?
The 3x doesn't hold: every tenfold increase in time cuts the multiplier by half a step
100 microseconds is the peak of the curve, not the whole curve. Between 100 microseconds and 1 second, the spec gives you a formula: allowed multiple = 4 − 0.2171 × ln(T), with T in microseconds. Run that formula and a very tidy pattern shows up — every order of magnitude longer in time, the allowed multiple drops by half a step:
| Time window | Multiple of rated power allowed |
|---|---|
| 100 μs | 3.0× |
| 1 ms | 2.5× |
| 10 ms | 2.0× |
| 100 ms | 1.5× |
| ≥ 1 s | 1.0× (rated) |
(The middle three rows are this post running the formula — the spec's own text only spells out the two endpoints, 100 μs and 1 second or longer. This fills in the gap.)
The curve is making an intuitive point: the shorter the spike, the more a component can ride it out on its own thermal mass, so the spec grants a higher multiple. The longer it drags on, the more that heat genuinely accumulates, and the allowed multiple keeps sliding back down to 1x — by the 1-second mark, the spec stops giving any slack at all. Your card either meets its rating or it doesn't.
That's the GPU side settled. What about the power supply — the part actually pushing the electrons? Is it asked to absorb even more?
What the power supply has to survive is, if anything, even more extreme
The spec has a matching curve for the power supply, and the requirement runs in the direction you'd expect — a PSU has to absorb not just what the GPU wants in that instant, but whatever every other component in the system is pulling at the same time. The ATX floor is: 200% within 100 microseconds, 180% within 1 millisecond, 160% within 10 milliseconds, 120% within 100 milliseconds.
| Time window | GPU-side limit | PSU-side limit |
|---|---|---|
| 100 μs | 300% | 200% |
| 1 ms | 250% | 180% |
| 10 ms | 200% | 160% |
| 100 ms | 150% | 120% |
Put the two curves side by side and you'll notice they weren't designed to line up — the PSU-side numbers are, if anything, more conservative than the GPU side. These are two independent floors; the spec never requires them to match, so there's no conspiracy to go looking for in the gap. It's just two separate test requirements written down in the same document.
One sentence covers both curves: for the GPU and the PSU alike, the shorter the spike, the more overload room the spec allows; stretch the window out and that room collapses fast, back toward the rated value. The PSU side of the ledger is settled — so what did that late-2023 spec revision actually change?
ATX 3.1 versus 3.0: what changed is how long the PSU rides out a power loss, not the spike multiplier
You might guess a spec revision would tighten the spike allowance a notch. It didn't, not at all. According to PSU maker Seasonic's own technical documentation, ATX 3.0 and 3.1 specify the exact same number for the instantaneous peak: 200% within 100 microseconds, unchanged — not loosened, not tightened.
What actually changed is a different spec called hold-up time — how long a PSU has to keep its output within spec, riding entirely on its internal capacitors, the instant after AC input is lost or glitches, buying the system time to react or shut down cleanly. ATX 3.0 required 17 milliseconds at full load; 3.1 lowered that floor to 12 milliseconds (while recommending 17 ms be maintained at 80% load). In other words, 3.1 actually relaxed the "how long can it ride out a power loss" requirement — it didn't tighten it.
You probably already know one of the other two changes: the connector's name changed from 12VHPWR to 12V-2x6, which is the mechanical-interlock revision covered in the first post in this series — it fixed whether the plug was fully seated. The other change gets mentioned less: the 12V voltage tolerance is identical across both versions, -7% to +5%. Not tightened, not loosened.
The connector got a new name, the interlock got fixed, hold-up time even got a bit more forgiving — all of that is written into the spec in black and white. But if you're thinking back to the case in the first post where a card spiked to 613 W, you probably already have a question forming.
The 613 W case doesn't fall under this rule at all: the spec permits "an instant," not "indefinitely"
Quick recap of the background: on 7 September 2026, an MSI RTX 5090 (rated 575 W from the factory) was running DLSS 5's new rendering feature when monitoring software logged board power at 613.5 W — and not as a spike that ticked up and settled back down, but sustained above 610 W the whole time. The first post already flagged this one clearly: it's a user report plus press coverage, with no vendor or independent lab confirmation.
But the point here isn't how much to trust that specific case. It's a more fundamental distinction: even if it's entirely accurate, it doesn't fall under the power excursion clause this post has been describing, at all. The reason is straightforward — an excursion is, by definition, a phenomenon measured in microseconds to seconds. The spec's own curve caps out and converges back to 1x at the 1-second mark, and there's no entry anywhere past that point that legalizes running "continuously" outside your rating. "Sustained above 610 W" describes a state lasting minutes, maybe longer — that's at least four or five orders of magnitude away from a 100-microsecond spike on the time axis. They are not the same kind of event.
Here's a way to picture the difference. Plenty of turbocharged performance cars have an "overboost" feature: floor the throttle for a few seconds to pass someone, and the turbo is allowed to briefly push past its normal boost setting, squeezing out extra power — and the system pulls it back down to the normal ceiling a few seconds later on its own, because the engine can't survive staying in overboost indefinitely. The spec permitting a GPU to spike to three times its rating is the same bet: short enough, and the system pulls it back down.
That distinction — an instantaneous spike versus a sustained overload — decides whether a software setting like "lower your power limit" can even reach the thing you're actually worried about. That's exactly what the next post takes apart: what lowering a power limit can and cannot actually do.
Frequently asked questions
Does your GPU spiking to 3x its rated power mean there's a design flaw in the connector?
No — it's not a defect, it's the power excursion clause written into Table 3-1 of the ATX12VO 2.01 spec: as long as the window is short enough (100 microseconds or less), a spike to 3x rated power is explicitly allowed, and the power supply side is required to survive a matching instantaneous peak (200% within 100 microseconds). This is a normal, spec-sanctioned transient, not an anomaly — you don't need to worry just because you've seen this number.
Is the 3x ceiling only good for that one 100-microsecond instant, with zero overload allowed after that?
No, it's not a cliff-edge. The spec uses a formula that lets the allowed multiple decay as the window lengthens: 3x at 100 microseconds, down to 2.5x at 1 millisecond, 2x at 10 milliseconds, 1.5x at 100 milliseconds, capping out at 1x (rated) by the 1-second mark. The longer the window, the less room your card legally has to exceed its rating.
Did ATX 3.1 tighten the instantaneous peak allowance compared to 3.0?
No. According to Seasonic's own technical documentation, both versions specify the same 200%-at-100-microseconds instantaneous peak for the power supply — neither loosened nor tightened. What actually changed was hold-up time (17 ms down to 12 ms at full load) and the connector's name (12VHPWR to 12V-2x6); the 12V voltage tolerance is identical across both versions.
Does the RTX 5090 case that hit 613 W on 7 September 2026 fall under what this excursion rule allows?
No. The excursion clause covers instantaneous spikes measured in microseconds to seconds. That case was monitoring software logging board power sustained above 610 W, not a spike that rose and settled back down. Nothing in the spec's excursion provisions covers running continuously above rated power — those are two entirely different things.
Where this stands, honestly
Every number in this post — the 3x, the 200%, hold-up time — comes from the spec itself or a PSU manufacturer's own technical documentation stating an allowed value or a floor. None of it is a measurement this post made itself. We don't have an oscilloscope, and there's no way for us to measure how high an actual GPU spikes within 100 microseconds — all this post can do is tell you what the spec permits, which is not the same as telling you what your specific card actually does. "The spec allows it" and "it actually happens" are two different statements.
This post originally wanted to cite something more concrete — a specific card, measured spiking to a wattage well past its rating within a very short window, so the abstract 3x number would have a real case to anchor to. That number got cut in the end: the source page failed to load on three separate attempts, leaving nothing but a headline with no way to verify the test conditions behind it. Not being able to find something is exactly that — this post says so plainly rather than dropping in a figure nobody here has actually checked.
If you're just here to find out whether your card is at risk: the spec-permitted instantaneous spike is a normal condition your card is designed to ride out, and this post alone isn't a reason to worry. What's actually worth paying attention to is the other situation covered in the next post — power draw sustained outside the rated envelope, not an instantaneous spike, which is the part software actually has some reach into.
The microsecond-scale power spikes this post covers are outside what PowerDoze touches, and outside what it's trying to touch — that's the spec's and the circuit board's job, and no piece of software reaches down to that timescale. What PowerDoze actually does is narrower: making sure your card doesn't end up like that 613 W case above, sustained outside its own rated envelope for the long haul. Want your GPU's power limit put back after a reboot by a power mode and an all-day, every-day schedule rule pointed at it, instead of retyping a command yourself every time?