NVIDIA GeForce RTX 5090 Founders Edition Hands On, Taken Apart

[mkppo]

We get a built in riser cable lol

I was hoping for the waterblocks with side terminals like the EK 3090 FE waterblock. Heatkiller made some for 4090 as well, here's hoping they make one for the FE or some AIB this round as well

 

[Vayra86]

Impressive. Now lets see it dissipate all those watts, we have seen various FEs that throttled. I am not saying this one will and if it indeed wont thats an engineering marvel.

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

 

[jesdals]

I would love this card if it wasnt for the price

 

[TheDeeGee]

Skip to thermal testing and notice the difference in comparison to 4090 FE. Then you'll probably understand why they designed 5090 they way they did.

Insane thermals on that prototype, holy moses!

 

[The Terrible Puddle]

Just buy a partner card if you want a waterblock for a convential PCB.
I am sure there will be lots of waterblocks for something like Asus cards.

 

[GhostRyder]

With precision measuring tools, CAD software, and CNC, it really wont be that difficult..

I don't disagree, I mostly meant though in respect to the fact the display outputs are on a separate piece as I figure it will have to incorporate that into the design and become the structure of the whole GPU.

 

[uplink777]

Hmm, more like 5090 CE = Cripple Edition

 

[TokyoQuaSaR]

Quite the contrary, a small PCB is desirable and an engineering feat that took many many years to achieve. One this small then is a marvel of technology. Smaller PCBs mean better signal integrity with as little loss as possible. So ridiculous isn't exactly the word I would use, it's closer to "remarkable"

Exactly my thoughts.

Well I wouldnt complain at least. And be honest, it wouldnt be a bad card. To me the founders edition was always the "reference" design with the cheapest pcb that everyone uses for waterblocks because it has the highest compatibility.

But go on, celebrate this developement further. F the enthousiasts!
The 7090 is just going to be a Geforce now Streaming device then. But I guess people will love it because it looks sleek. You will own nothing and be happy.

The objective is to make a nice cooling solution, and I do think it's an awesome way to do it. The other way would be water cooling, but as you say the FE is supposed to be cheaper than that.

 

[londiste]

The flow through design doesn’t help THAT much, it’s in the realm of 5c. Just the overall mass loss of heatsink is questionable. I have a feeling FE will either run well below 575w in stock/normal conditions, or AIB cards will really be able to stretch their legs under water.

But 5C is quite significant in this context, no?

Sure, though for the breakaway boards im not so sure ribbon cables are exactly "clean"

I do have some concerns about how well or reliably the ribbon cables and connectors work with PCIe5 or some faster display outputs but in context of the main PCB itself these ribbon cables might be a blessing in disguise. Instead of routing the tricky traces to the PCIe fingers or display outputs at the edge of PCB somewhere these are taken off the PCB in some convenient point where they have as little effect as possible on everything else.

 

[rv8000]

But 5C is quite significant in this context, no?

I do have some concerns about how well or reliably the ribbon cables and connectors work with PCIe5 or some faster display outputs but in context of the main PCB itself these ribbon cables might be a blessing in disguise. Instead of routing the tricky traces to the PCIe fingers or display outputs at the edge of PCB somewhere these are taken off the PCB in some convenient point where they have as little effect as possible on everything else.

Yes, but at the same time the overall mass loss of heatsink is probably going to have a much larger impact, we also don’t know if a second flow through area will give the same amount of benefit; we didn’t even get a good look at the bare cooler, daughter boards, and connectors.

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.

 

[TheinsanegamerN]

At this point, we just need a socket on the motherboard for the GPU.

As neat as that is, the limitations would become immediately apparent, sinc eGPu generations are pushing boundaries and I cannot imagine the memory issues. There's a reasont he memory chips are crammed so close to the GPU, and memory standards change so frequently, that socketing the GPU would accomplish nothing for longevity.

Three separate PCBs connected by ribbon cables is pushing the limits of keeping the current paradigm and could lead to all sorts of problems if the cables become disconnected.

Cables are not an issue. If they become a problem, re attatch them. There will likely be a dollop of hot glue on them to prevent them from moving anyway.

 

[londiste]

Yes, but at the same time the overall mass loss of heatsink is probably going to have a much larger impact, we also don’t know if a second flow through area will give the same amount of benefit; we didn’t even get a good look at the bare cooler, daughter boards, and connectors.

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.

Heatsink mass is not really that big of a factor. Radiators-heatsinks are all about area and temperature difference. Flow should be positive for the temperature delta.
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.

 

[rv8000]

Heatsink mass is not really that big of a factor. Radiators-heatsinks are all about area and temperature difference. Flow should be positive for the temperature delta.
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.

If you’re reducing the mass via cutting the height of fins (thickness of the card), you are reducing the surface area, that’s what I was getting at (wrong terminology). Coming from the 3+ slot design of the 4090 fe to the 2slot, there are going to be minimal surface area differences that are probably in favor of the 4090 fe cooler. Add 100w+ or more and that’s not an ideal recipe.

Will be neat to see performance differences on water.

 

[Suspecto]

I don't think I've seen worse design than this, remind me of Apple products, everything is cramped together and exposed to heat and if something breaks or shorts, explodes, it is gg.

Shit product for the premium price.

 

[Darmok N Jalad]

Rather than seeing the design as a negative, it actually lends itself better to repairability. If the ports are on a daughter card, that means they could more easily be swapped out in the event of a failure. Back to the 2013 Mac Pro, it, too, uses ribbon cabled attached to multiple cards. I own one made back in 2014 at the latest, and if anything those connections are hard to break free. If it’s designed well, it should be a good product.

 

[Lycanwolfen]

I'll wait till I see real world review, tear down and other stuff.

 

[Solaris17]

Imagine what they have or will have in the wild shortly..

We already do. We fit racks full of them. Here is Blackwell specifically B200. This is a servethehome pic. Our HGX boards do 8 per tray and are black.

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.

 

[freeagent]

We already do. We fit racks full of them. Here is Blackwell specifically B200. This is a servethehome pic. Our HGX boards do 8 per tray and are black.

View attachment 379328


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.

You are a lucky fellow

 

[igormp]

We already do. We fit racks full of them. Here is Blackwell specifically B200. This is a servethehome pic. Our HGX boards do 8 per tray and are black.

View attachment 379328


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.

Just to make it clear for others, that's an SXM socket, Nvidia has been using those for their highest end x100 GPUs for quite some time (think P100, V100, A100, H200, etc etc). Easy way to have nvlink and power delivery without a cabling mess.

You are a lucky fellow

I guess that's one of the perks of working in a big tech lol
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