If you've read the rest of this series, there's a plan probably forming in your head: lower the power limit a bit, and doesn't that ease every margin problem covered so far? Current scales down linearly with wattage, so the safety margin should widen on its own — the logic sounds solid, and you may already have a number in mind.
In February 2026, a card that had already done exactly that — and gone further than most people would — melted anyway. That GIGABYTE AORUS RTX 5090 MASTER ICE was undervolted to 0.9V, with its power limit capped at 500W (down from a 600W stock limit). Both the GPU side and the cable side came away with damage. TweakTown, Guru3D and HotHardware all covered it.
This post pulls apart what lowering a power limit can actually do from what it can't — including the biggest gap this post ran into while researching it: after searching everything available, there is no first-party dataset anywhere of the same card, tested at different power limits, with measured connector temperature and per-pin current. Every claim that lowering the limit helps — including the table later in this post — is a theoretical extrapolation, not a measurement. That needs to be said up front, so nothing below gets mistaken for a recommendation.
Start with the card that was already at 500W — what does it actually prove
You may have already seen this card in the first post's timeline table. Here are the details: in February 2026, a GIGABYTE AORUS RTX 5090 MASTER ICE had already been undervolted to 0.9V and power-limited to 500W — both levers pulled, both set well below stock. The result was scorch damage on both the GPU side and the cable side.
This one card doesn't prove "lowering the power limit does nothing" — a single case can't overturn a probabilistic claim, any more than one person getting hurt while wearing a seatbelt proves seatbelts don't help anyone. But it does overturn a stronger claim: "lowering the power limit means it won't melt." Disproving an absolute claim only takes one counter-example, and this card is that counter-example.
That doesn't mean lowering the limit does nothing — it genuinely accomplishes one thing. It just isn't the thing you should actually be worried about.
What lowering the limit actually does: it brings down the average current on all six wires together
This part isn't up for debate — it's Ohm's law. Watts divided by volts equals amps; the voltage is fixed at 12V, so however much you cut the wattage, the average current split across the six wires drops by the same proportion.
| Power limit | Average current per wire | % of the 9.2A rating used | Margin left |
|---|---|---|---|
| 600W (stock) | ~8.3A | ~90% | ~10% |
| 500W | ~6.9A | ~75% | ~25% |
| 400W | ~5.6A | ~60% | ~40% |
Drop the limit from 600W to 500W and the margin goes from 10% to roughly a quarter. Drop it to 400W and the margin approaches 40%. That's real, and the melted card above genuinely had a wider margin than stock — it went from 600W's 10% margin to 500W's 25%, and it still melted. The margin really did widen. But do the actual root causes covered in earlier posts in this series improve along with it?
Four things lowering the limit does not touch, at all
| Problem | Does lowering the limit fix it | Why |
|---|---|---|
| Current imbalance (22A on one wire, 0 on the rest) | No | The split ratio between wires stays the same — only the absolute totals shrink |
| Not fully seated, or lateral strain deforming the terminals | No | This is a mechanical contact problem, unrelated to total wattage |
| Abnormal contact resistance in the cable or terminal | No | It only shrinks total heat output — it doesn't change where the heat concentrates |
| Microsecond-scale power excursions (spec-legal spikes to 3x rated) | Barely | A power limit governs an average — it has no reach into microsecond events |
| How far average per-pin current sits from its rating | Yes | This scales directly with wattage, as the table above confirms |
The first row is the most counterintuitive one here, and it's worth sitting with for a second.
The fourth row is worth pausing on too, because it connects directly to the previous post. The spec legally allows a GPU to spike to three times its rating within 100 microseconds, while a power limit governs an average sustained over seconds or longer — the two operate four or five orders of magnitude apart on the time axis.
Looking at that table, you might already be asking the more fundamental question — who actually measured any of this "can" and "can't"?
This is the biggest gap this whole series has run into
Time for a plain statement, up front, not saved for the closing honesty section: the table above is a linear extrapolation from Ohm's law, and that part isn't in dispute. But behind the entire narrative that lowering the power limit makes things safer, after searching everything available, there is no first-party data anywhere showing the same card, set to 100%, 80% and 70% power limits, with measured connector temperature and per-pin current. No vendor white paper has run that comparison. No independent outlet has run that experiment. Not one.
Which means every claim in this post — in this whole series — that sounds like "lower power equals safer" is currently a theoretical extrapolation, not something measured and confirmed. This isn't something this post is hiding. It's something decided up front, before writing a word of it, to say plainly.
Searching every technical opinion available, not one independent tester recommends this
| Source | Position |
|---|---|
| der8auer | The safety factor is only 1.1, and NVIDIA never implemented current balancing — the problem is in the design, not in how you set your power limit |
| Buildzoid | Add back multiple independent shunt resistors, so the card can actually see what each wire is carrying |
| Igor's Lab | The hardware side should either do active load balancing, or at minimum monitor for anomalies and cut power outright |
| NVIDIA (official) | Has never recommended lowering power limits — the sole official guidance is to ensure the connector is fully seated |
| Board partner fixes | Zotac (lights up once seated), ASUS (per-pin sensing with a warning at 9.2A), ASRock (overheat protection) — all hardware fixes, none of them telling you to change a software setting |
Igor's Lab also put together a power-tier framework, and it's worth being precise about what it does and doesn't say: it comes from a discussion of whether board designs need active load balancing or whether monitoring alone is enough — a question about how partners should design their circuits. It never treats "you lowering your own power limit" as an option within that framework. The table below is this post's extension of that framework, not Igor's own recommendation, and that distinction matters.
| Power range | Igor's assessment |
|---|---|
| Under 350W | Balancing is a monitoring-only concern |
| 350–500W | Active regulation is "increasingly useful" |
| Above 500W | Active load sharing is "a de facto industry standard" |
Converted to current, the "above 500W" tier works out to roughly 6.9 to 8.3 amps per wire — which happens to be exactly the range you'd land in by setting a power limit anywhere between 500 and 600W, per the table earlier in this post. Nobody recommends lowering the limit as the fix — so if you're still going to do it anyway, what does it cost you?
The cost: the RTX 5090 pays a much steeper performance penalty than the RTX 4090
Start with the RTX 4090. Per Tom's Hardware's October 2022 testing, capping the power limit at 80% cost roughly 3% of performance; at 70%, one test set still retained 97.8% (a 2.2% loss); at 60%, five AAA titles at 4K averaged an 8% loss. These three numbers don't sit on one smooth curve — they likely came from different test batches with different games or resolutions, and this post isn't going to force them into a tidy curve that isn't really there. But the direction is clear: as long as the limit isn't cut in half, the loss stays in the low single digits; it only starts to feel real once you're below 60%.
The RTX 5090 pays a much steeper price. The same source shows: at a 400W limit (roughly 67% of the 600W stock rating), performance comes out to about 83% of what the card does at a full 600W — put another way, 600W burns 50% more power than 400W for only a 20.3% performance gain. Flip it around: cut the power by a third, and you keep 83% of your performance, a 17% loss.
| Cut to roughly two-thirds power limit | Performance kept | Performance lost |
|---|---|---|
| RTX 4090 (70%) | 97.8% | ~2.2% |
| RTX 5090 (67% / 400W) | ~83% | ~17% |
Cut both cards to roughly the same two-thirds mark, and the RTX 4090 barely registers the difference while the RTX 5090 is one you'll actually feel while playing. A caveat that needs to sit right next to this number: the original source for the 5090 figures is TechPowerUp, and three separate attempts to reach that page returned a 403 — this post can only cite the number as relayed by other outlets, with no way to verify the resolution, game list, or other test conditions behind it. Treat it as secondhand reporting, not something this post checked against the primary source itself.
Worth separating two different levers here: lowering a power limit and undervolting are not the same knob. ASUS's official line on undervolting the 5090 is that average frame rates typically differ by only one to three frames, while system power draw drops by 70 to 170W — a far gentler trade-off than cutting the power limit outright. But don't forget the card that opened this post: it was undervolted to 0.9V and power-limited to 500W at the same time, both levers pulled as far as they'd go, and it melted anyway. Undervolting being gentler on performance doesn't mean it's more effective against melting.
With the cost accounted for, how should you actually think about all of this?
Frequently asked questions
Does lowering your GPU's power limit prevent connector melting?
No — there's a clear counter-example. In February 2026, a GIGABYTE AORUS RTX 5090 MASTER ICE already undervolted to 0.9V with its power limit capped at 500W (down from 600W stock) melted anyway, with damage on both the GPU side and the cable side, covered by TweakTown, Guru3D and HotHardware. Lowering the limit does bring down average current per wire, but it has no effect on current imbalance, poor contact, or microsecond-scale spikes — the actual root causes.
What does lowering your power limit actually change?
Just one thing: it scales the average current on all six wires down linearly with wattage, widening a margin that starts out very thin. At 600W each wire carries about 8.3A (90% of the 9.2A rating); at 500W, about 6.9A (75%); at 400W, about 5.6A (60%). The margin genuinely widens — that doesn't mean the problem is solved.
Is there measured data showing that a lower power limit reduces connector temperature or improves current sharing?
No. This is the biggest gap this post ran into: there is no first-party data anywhere of the same card tested at different power limits with measured connector temperature and per-pin current. Every claim supporting lower power limits, including the table in this post, is a theoretical extrapolation from Ohm's law, not a measurement.
Does any expert recommend lowering the power limit to reduce melting risk?
No independent tester found during research recommends this. der8auer points to a safety factor of only 1.1 and NVIDIA's lack of current balancing; Buildzoid argues for adding back multiple shunt resistors; Igor's Lab argues for active hardware-level load balancing or a hard monitoring cutoff. NVIDIA and board partners have never recommended lowering power limits either — the official guidance is always to ensure the connector is fully seated, and every partner fix is a hardware one.
Does lowering the power limit hurt your gaming performance much?
Depends how far, and which card. An RTX 4090 capped at 70% still retained 97.8% of its performance in one test; at 60%, five AAA titles at 4K averaged an 8% loss. The RTX 5090 pays a noticeably higher price: capped to roughly 67% (400W), performance drops to about 83% of its 600W figure, a 17% loss. That 5090 figure's original source page failed to load after repeated attempts, so it should be cited as secondhand reporting, not independently verified.
Where this stands, honestly
Tying this together, the honest conclusion looks like this: lowering the power limit isn't useless — it genuinely brings down the average current on all six wires, and that isn't in question. But what it fixes is "what percentage of the rating you're using," not "why connectors melt" — current imbalance, poor seating, abnormal contact resistance, and microsecond spikes are the actual root causes on record, and lowering the power limit doesn't touch a single one of them. We searched everything available and found no independent tester recommending this as mitigation; NVIDIA and board partners never have either — every recommendation and every fix from them lives on the hardware side.
If you opened this post hoping to confirm that "lowering the power limit means it won't melt," this post can't give you that guarantee — that February 2026 5090, already running both levers at once, has already disproven it. This is a case where you don't need us, or any power-management software at all. What actually matters happens at the hardware level: confirm the connector is fully seated and you heard it click, avoid putting lateral strain on that cable, consider a card with per-pin sensing if this matters enough to you (currently only the ASUS ROG Astral offers it), or contact your board partner directly.
What PowerDoze does in all of this is much narrower, but it's something it can actually deliver: it re-applies your chosen power limit for you, because nvidia-smi -pl itself is a runtime setting that the driver resets on every restart. That happens on a mode switch, and at startup if a schedule rule pointing at that mode is in effect right then — so an every-day, all-day rule covers every reboot. This post has the full walkthrough. That solves the situation from the first post in this series — you changing nothing while a software update pushed a card to 613W — by keeping your card running inside its own rated envelope instead of quietly drifting outside it. It will not balance current between wires, it will not fix contact resistance, and it will not make a connector any less likely to melt — this post has already explained why no software can do that.
Want your power limit to still be set after a reboot, without retyping a command yourself every time?