Thursday, January 9th 2025
NVIDIA GeForce RTX 5090 Founders Edition Hands On, Taken Apart
At the 2025 International CES, we went hands on with the new NVIDIA GeForce RTX 5090 Founders Edition "Blackwell" graphics card. This thing is huge—longer and taller than the RTX 4090 FE, and yet just 2 slots thick. This is because NVIDIA's designers have figured out that the heat dissipation area of the heatsink lost to thinning the card can be recovered by stretching it in other directions. The card retains the essential aesthetic of Founders Edition cards from the past two generations going back to the RTX 30-series, but changes the concept of the dual-axial flow-through.
While past generations used an intake fan on one side, blowing onto the PCB, and another fan at the tail end of the backplate pull air through the heatsink and out the back, the RTX 5090 FE has two large fans, both of which blow cold air through the heatsink, and out the back of the card. The PCB is located in the center of the card, and relies on a set of breakout PCBs for host interface and display outputs.
The biggest component on the PCB, which takes up nearly 1/3 of its board area, is the "GB202" GPU on which the RTX 5090 is based. The GPU has a gigantic pin-count not just for its power needs, but its 512-bit GDDR7 memory interface. At one corner of the PCB, which sticks out from the top of the cooler, is the card's 12V2x6 power input, which is rated for 600 W (we don't know the TGP of the RTX 5090 yet).
The card uses a VRM solution with 19 phases for the VGPU, and 8 phases for the memory. Much like an AI GPU board, NVIDIA resorted to high-density PCB engineering, not wasting any space on either sides of the PCB. The chokes and DRMOS (made by MPS) are on the obverse side, surrounding the GPU on three sides; and the capacitors are on the reverse side.
The reverse side of the PCB has connectors that lead to its two breakout components. The first one connects to a PCB with the PCI-Express 5.0 x16 gold fingers. The other connector leads to the display I/O breakout. Both connections are made using thin ribbon cables like the ones you find in laptops, and routed along the edges of the cooler, so as not to impede airflow.
While past generations used an intake fan on one side, blowing onto the PCB, and another fan at the tail end of the backplate pull air through the heatsink and out the back, the RTX 5090 FE has two large fans, both of which blow cold air through the heatsink, and out the back of the card. The PCB is located in the center of the card, and relies on a set of breakout PCBs for host interface and display outputs.
The card uses a VRM solution with 19 phases for the VGPU, and 8 phases for the memory. Much like an AI GPU board, NVIDIA resorted to high-density PCB engineering, not wasting any space on either sides of the PCB. The chokes and DRMOS (made by MPS) are on the obverse side, surrounding the GPU on three sides; and the capacitors are on the reverse side.
The reverse side of the PCB has connectors that lead to its two breakout components. The first one connects to a PCB with the PCI-Express 5.0 x16 gold fingers. The other connector leads to the display I/O breakout. Both connections are made using thin ribbon cables like the ones you find in laptops, and routed along the edges of the cooler, so as not to impede airflow.
93 Comments on NVIDIA GeForce RTX 5090 Founders Edition Hands On, Taken Apart
I think they moved towards the wind tunnel idea of rack mounted cooling here. Just force air through. I hope it wont also have a similar noise signature :)
I am sure there will be lots of waterblocks for something like Asus cards.
Their FE coolers do a pretty good job but smaller cooler + significantly higher tdp/tgp isn’t a recipe for cooler or the same temps. Nvidia and AIBs have already stretched air cooling to it’s limits, at some point you’re just not going to overcome science when adding another 100-150w.
I guess we will see what the cooler itself actually looks like compared to 4090 one but with less PCB in the way and the claim of 5090 being slightly longer and wider the area might even be bigger.
Will be neat to see performance differences on water.
Shit product for the premium price.
They looks integrated, but they arent. They are screwed down. and a massive heatsink is put them and screwed in again. The HGX board itself just contains retimers and the primary power delivery circuitry basically taking say a 8pin molex and running it through the copper layer to the GPU socket. All the power delivery, vram and GPU core are on those cards you see. They fit in the palm of my hand. Previous architectures and future architectures also come in similar formfactors using gen 2 HGX boards etc.
Its just now coming to the consumer space, but it has been in this similar form factor for years.
The company I work for is barely starting to work on managing our own stack instead of using managed solutions, hopefully one day having our stuff on-prem will make enough sense so I can toy with those things as well haha