A Galax RTX 5090 that left the factory with two 16-pin sockets spent an hour on a livestream pulling up to 900W through three ordinary 8-pin PCIe connectors, with the wires peaking at 35 degrees C. Nothing in the card's firmware knew the difference. The question behind the build is whether the connector the industry moved away from can still feed a 575W card.
Three 8-pins, 900 watts, 35 degrees
Brazilian YouTube channel TecLab took a Galax HOF OC LAB XOC, a top-bin board that ships with two 16-pin connectors, and stripped it to the bare PCB. Three 8-pin sockets went on in their place, the on-board sense pins were rewired so the card would report full power available rather than derate on a handshake that no longer matched the hardware, and a custom die-cooling setup went on top. The changes stopped at the hardware layer, with no firmware modification required.
Over the stream the team walked the power draw from 400W up to 900W. At the top the card pulled around 120A at 3,400 MHz, and wire temperatures peaked near 35 degrees C. Dropping to two cables put it at 66A and 3,200 MHz. A single cable still held 65A and 3,100 MHz, with the wires reaching a warmer 65 degrees C. That last pair is the odd one: current and clocks barely moved between two cables and one, which says the card was being held back by something other than the cabling, and the stream does not identify what. These are TecLab's own readings from a live session rather than an independent bench, and they are worth reading that way.
Why 150W is not the conductor's limit
An 8-pin PCIe connector is rated for 150W, which is what makes the result look impossible at first glance. That figure is a specification with a wide safety margin built into it, not a measurement of what the metal will carry. The modified card ran nine +12V rails at roughly 8 to 8.5A each, which is ordinary Mini-Fit Jr terminal territory and multiplies out to something close to the 900W the card actually drew. TecLab started the experiment with XT90 connectors rated at 90A apiece before concluding that three 8-pins would do the job.
This does not turn a triple-8-pin RTX 5090 into a product. It will not go in a normal case, and it exists because somebody was willing to solder onto a flagship board. What it does is put numbers against a complaint that has followed the 16-pin connector since launch, which is that the failures concentrate current across a handful of small contacts. The run showed none of the current and temperature deviations a failing 16-pin cable gives off, and melted plugs on cards and power supplies are still being reported. Days earlier TecLab had run power wiring directly onto a card's PCB and called that a definitive end to melting, while accepting it is not something a gamer can do at home.
Someone finally built the 16GB RTX 3070 Ti
NVIDIA shipped the RTX 3070 Ti with 8GB of GDDR6X. A 16GB version got far enough to appear in driver identifiers and to surface as an engineering sample with 512 GB/s of bandwidth, and it never reached retail. Brazilian YouTuber fmklab has now assembled a working one, and by his account it is the first publicly documented rebuild that runs.
The base is an ASUS TUF RTX 3070 board rather than a 3070 Ti, because that PCB is already laid out for plain GDDR6 and the double-capacity GDDR6X modules a real 3070 Ti would need were far more expensive. The GA104 came off and was replaced with the fully enabled part from a 3070 Ti, and the eight Samsung 8Gb chips were swapped for 16Gb ones. That alone was not enough: GPU-Z still read 8GB after the first boot, and the memory straps had to be rewritten, worked out with NVIDIA BIOS Reader, before the full pool appeared, according to Tom's Hardware's account of the rebuild.
Validation went further than a screenshot. NVIDIA's MATS diagnostic reported no memory errors, a local language model filled nearly the whole 16GB, Call of Duty: Warzone used more than 13GB, and Unigine Superposition returned 5,367 points. The capacity cost something, because 16 Gbps GDDR6 on the 256-bit bus gives 512 GB/s against the 608 GB/s the retail card gets from 19 Gbps GDDR6X. Doubling the pool does not make the card faster on its own and only pays back in workloads that were running out of VRAM rather than out of throughput.
It is not an upgrade path. Two donor cards, BGA work on both the die and the memory, and a strap edit buy a capacity an RTX 5060 Ti 16GB matches with a warranty attached. What the build settles is the older question of whether that cancelled configuration ever worked, and it does.