@@ExtremeGamersBenchmarks 2 slot, sff ready, design. although we'll have to wait until benchmarks come out to really see if its as good as normal air cooled designs
Not the insane 512 bit memory bus that was last used 16 years ago by GTX 285 because it's so difficult to implement? And the fact that they managed to fit it onto a small PCB too...
@@ExtremeGamersBenchmarks lets not forget that its also a 600W, so the fact that its just a 2 slot gpu for such power, the cooling is impressive for its size, hopefully its not a bottleneck
Thinking like an engineer, you may not need 58 different tools. If all the bends have the same radius, they could just use the same tool, but stamp the different parts a different number of times for each piece. You can calculate the deformation at a given pressure for the fin material, so for example, the least bent fin gets 1 stamp press, and the most gets n+58 stamps. The process becomes longer and more involved, but the cost of the tooling could remain within reason. Great video, very interesting, I always love your work!
Thank you. That's exactly what I immediately thought looking at those fins. Nobody in their right mind would make 58 different tools just for a single radiator, instead the shape of those fins look like something you can make with one customizable tool. Which is still pretty cool and much more cost efficient. It might work just as you described or it might have different customizable pressure parameters for each desired deformation of a fin and just cycles through them.
Could they also be stacking up the fin plates on top of each other and then stamping them all at once? The top stamp would be the sharpest, and all others after that would be greater radius as they all mold around each other in the press.
I was about to mention, PCWorld's "Nvidia Talks RTX 5090 Founders Edition Design" explains the concave 'grill' shape, the fin variations, and the triple PCBs. And, indeed, Roman pretty well nailed them all. Other items in that video are the purpose of the side vent, which is not just another exhaust point, as well as fan blade design, and possibly a couple other elements I'm forgetting.
Only reason I can think of for someone to a) employ an ex EK waterblock designer and b) get them to do an accurate CAD model of 5090... is to make future 5090 waterblocks
The waterblock manufacturers have long had the AIB's board designs from the AIB's with the full knowledge of Nvidia. They are likely to be available on launch day.
I can think of another reason - because 3D design skills are transferable and there was a sudden oversupply so to speak of ex EK employees looking for work that the tech RUclips community specifically went out of their way to find jobs for
PCWorld has an interview with a NVIDIA product manager yesterday which confirms many of the assertions raised here; the fins are concave to optimize for the pressure delta of the passthrough fans, the angled fins are to redirect air to make a "shield" to deter cannibalistic looping air. Definitely worth a watch.
Tell me Joe is doing all that work because there is a Grizzly FE block in the works , don't waste this man talent . By the way we need an entire WC lineup from Grizzly !
He better get used to it. AI isn't going anywhere, we're entering a new phase of graphics rendering and generation. If we want Hollywood grade movie CGI running in real time on a single PC, we're going to have to rely on AI. We're nowhere close to be able to brute force those level of graphics in real time.
@@03chrisv Sorry, but modern day graphics can only ever be called brute forcing compared to 8th generation consoles. Ryse Son of Rome anyone? God of war on Ps4? Absolutely gorgeous games that run amazing compared to modern day SLOP.
RTX 5090: 30% faster than 4090 (ignoring framegen, pure raster) 30% more power consumption than 4090 30% more expensive 30% larger die size on the same process node Such innovation, much wow.
Crazy is the 2080 ti with dedicated Tensor cores vs 1080ti had a higher delta gain. Hey at least we can rely on smoke and mirrors to make up for the marketing and hype shortfalls. :kookoo:
@@aberkae the 2080ti had about a 25% delta in raster, that's not higher than 30%. The msrp from 1080ti was $700 and 2080ti was $1000 usd, that's 30%. so we got 25% more performance for 30% more money back then. The 5090 has a higher delta gain to the 4090
As someone who works in a metalshop with laser and punching machines, what they most likely did to get the decreasing profile in the fins, is use the same tool but use less force while punching in the profile. It's possible they could use some Ertalon (nylon) in the die to have some support when the profile gets smaller. We use wheel forges in CNC punching machines that roll over the plate to make shapes like that (but much longer) and then lower the pressure so it will flare out the profile into the metal sheet and makes it look better.
Usually the dimples, bends or louvers would be to increase turbulence and thus increase the air side heat transfer coefficient. The concavity of the fins may be to try and increase flow velocity through the center by increasing the resistance on the edges (since there is more fin surface area). Honestly the thermal design of this card looks amazing and a real treat for thermal engineers. I hope I can get my hands on one to actually see it in person. I wonder if they have brazed the heat pipes to the vapor chamber or they have them directly connected to the chamber. Can't wait for you or Steve to cut them open!
I think dimples serve to decrease turbulence, because why would you want turbulence. I know from practice when i put 4090s on a rig 20cm away from each other they were overheating and when i put my hand between them there was huge turbulence. Air was not going the way it was intended to. It seemed like fans were as much sucking as blowing the air. It makes me think that if you want to make a PC rig that has exceptional thermals you better do it flawlessly or give up all altogether. When you have weak airflow atleast you didn't create turbulence.
I bet you the primary reason for the dimples is optimizing air flow and reducing pressure when multiple cards are stacked together. Remember, the 5090 is only a gaming card in name at this point. A lot of these will end up getting tossed into AI farms. The partner cards will likely get stripped again like what happened with the 4090. That's why most the cards on ebay are listed as "No core or VRAM". Jensen knows exactly what he's doing.
I think they said something with 3d vapor chamber my guess is like you said The vapor chamber hugs the heatpipes for more surface area and is probably brazed as well Since looking at the design they took a WHOLE HECK load of effort into making it a 2 slot card
3rd parties have to justify their ever-increasing premiums. The only thing they have left to market to rational customers outside of mild factory OCs are quieter and cooler GPUs
@@auturgicflosculator2183They look really good to me. This FE just looks boring. Also, AIBs improved greatly along with FE for years. I am sure that more premium models like astral, master will be much colder than this FE.
15:30 Nvidia can use only 1 stamping tool, Basically stacking the sheets and stamping with a curvature value. The highest sheet will receive the most curvature while the lowest in the stack will be minimal.
regarding the stamping tool, I doubt that they made 52 different stamping tools, I would say it's more likely that it is stamped with varying pressure with the same tool to get the different bends. Still an expensive process as they need to keep all the plates organized for assembly though, even if they only used one tool.
@@asmi06 Was thinking about this too, perhaps there's a manufacturing technique which allows for stacking multiple plates at the same time, reducing the cost of producing parts like this significantly?
They would have needed 58 tools because you also have the feature on the top for the spherical dent where each sheet would have a different radius. Also if you use the same tool with varying pressure your thickness in the bend would change. 58 tools at 5K each is only 290k, it is a drop in the bucket, the real cost would be in assembly, where you would need a machine that automatically stacks all 58 sheets in the right order before pressing the heat pipes in. They already have a couple stamps for the triangle portion so it's no big deal. The cost is also divided across 4 models on the consumer side, and may also carry over to their B series pro cards in some capacity. Looks like NVIDIA spent more time and money designing this cooler than AMD did on their RDNA 4 architecture, or Intel on Arrow Lake.
@@asmi06 I've actually done this in a press shop. Each section of the larger die would have its own specific section designed to make one fin in an assembly like this one. You simply increase the surface area of the entire die and use a larger machine. Individual fins could be cut from the resulting sheet afterwards easily. These fins are quite thin as well, making everything operate on a smaller scale of machinery needed which would lower the cost of tooling investment.
Regarding the shape of each fin being slighty different: Perhaps the method used is to stack the fins and stamp them together in a single step, with a single two-part mold. The added thickness of each fin would then ensure each one is a little bit different. Think of it as if the stamping mold is for doing a "D" shaped curve, and the width/thickness of the "D" itself is given by the added fins. The keynote's images seem to confirm this method. (maybe some sort of liquid is sprayed between the fins first to allow easy separation afterwards, but that is a detail of the process anyway). If this does not aling exactly with what we see in the images, maybe there is some kind of sacrificial separator between each fin, or those are the fins that go to the other side of the cooler...
PC World did a video with 'Justin Walker - Senior Director Of Products' where they went over the card and explained about the PCIE board, where it slots in, and also the PCB of the IO and how and where that fits into the card. The two slots at the top are for airflow, it's angled to blow air in a way so that air cannot wash back around and get sucked back into the fans, effectively creating an air barrier. And yes you are correct, the fins are concaved for pressure reasons.
As an engineer that specializes in cooling, the fin designs are likely serving a few purposes. Varying the length of the fins with the concave shape to the overall fin stack allows them to control heat transfer to the fins from the heat pipes in a very efficient way. You would want less heat transferred to the area above the fan hub for example because there is significantly less flow over those fins. This is probably the same reason for the increasingly round indent in the fins toward the center of the card, except in that case they are using the fin shape to slow down the flow and add more surface area for heat transfer to take place. Just my 2 cents. Great video!
@@deejnutz2068 well, in my case with the 5800x3D and an Rtx 4070 the limiting factor in all games I play is always my graphics card. I play everything in 4K bc I’m a couch gamer. The 4070 does pretty good for most of my games “most games I can do 4K 120hz on medium-high settings” but I’ll likely get a 5080 if I can get one for msrp. Also I never see my cpu usage go above 20-30% usage. So genuinely no reason to upgrade to am5 yet.
I also downloaded the SVG chart from NVIDIA and broke it down to a pixel perfect count. My result showed a 27.56% uplift over the 4090 in the Far Cry 6 benchmark.
15:15 the different curve in each fin is possible with a single press. In a stack of fins, the closest to the press will receive the most deflection and it tapers off exactly as you see it on the card, due to each fin absorbing some deflection by compressing.
You don't need different tools for the differently bent fins, just press the tool deeper for each progressively further bent fin. Any CNC stamping tool could do this.
Every now and then, I feel then urge to pop on here and thank you, once again, for the time and work you put in to produce both DE and EN videos. You could have taken shortcuts to accomplish this or chose one language over the other to the dismay of your audience, but you chose the quality path and we appreciate you for it!
For the 58 different fin designs, it could be that they designed a wide stamping tool with an extrusion for those curved bits that goes from a larger/deeper extrusion on one side, and slopes up to a shallower extrusion on the other. Load up a wide (fin width x 58) sheet of metal in there, and stamp out ALL 58 fins in one go. Once stamped, the sheet would get machined apart into the 58 separate fins, and that's that. This would result in fins whose extrusion depth changes across each individual fin's width, but if all fins change similarly, then that shouldn't be a problem.
5:59 frame generation is not more performance. Better frame smoothness, but system latency is even higher which goes counter to the conventional measurement of performance where higher FPS means better responsiveness.
Fake frames are analogous to an automatic transmission if you think about it. Automatic transmission can't anticipate the road, just like how the gamer will change the scene. It's an instantaneous smoothing.
And all the fastest cars are? Automatic/electronic shifting, and what do they all have? Assistance, ie traction control, differentials, abs, and so on. Nvidia might be adding latency but I assume their but unveil of reflex 2 is to counter this, much of what they spoke about was the use of ai to render frames in a more accurate light to how it should be. Comparing a car to a computer is already a stretch and it's quite a pointless analogy, nobody needed an analogy. And you can't actually state it has higher latency without seeing it run, my guess is DLSS 4 limits the amount of native frames slightly, so rather than struggle to render 40 it does 30 and reduces latency through this and reflex 2, Nvidia may be greedy and may be cornering the market, but they aren't stupid, DLSS, even 1.0 was quite a piece of software, and they rarely mess up, I highly doubt that they won't deliver on this occasion. Most people moaning about latency and don't even benefit from 5ms, let alone the 20s they so badly dread, if it was poor, only a small portion would truly be affected in any forms meaningful.
From what I've seen the different bent radiuses indicate that they don't cut the fins from single sheet but rather cut them in one big stack. This is why the radius changes. Each fin has increased radius by the thickness of the fin.
I mean...Nvidia is the second-most valuable corporation in the entire world, I'm pretty sure they can afford a few dozen or even a few thousand stamping tools if they want. That's not even lunch money to them :) Anyway, I really just wanted to say that I always appreciate the depth and detail of your videos, and the fact that you shoot every video twice. I especially appreciate how you bring your knowledge and experience of the manufacturing and retail aspects of production and products into your videos - these are things that most other content creators can only speculate on, whereas you bring years and years of first-hand experience THANK YOU!!!
Impressive! I would love to see if you can technically dissipate 575watt of heat with two fans by just heating up a block with the same amount of heat pipes and two 120 fans to do a proof of concept. I know the fins would be impossible to reproduce but that should give us an idea if it’s really possible that nvidia might actually be onto something or they are compromising performance to meet their SFX requirements.
I think it's possible and even likely that they made a stamping tool with 57 different shims that can be placed over it in order to get 58 different depths of the curve. stamp 1000 or so at one depth, add a shim, stamp 1000 more at the next depth, and so on.
Regarding the 5070 = 4090 marketing nonsense: according to Nvidia's own footnotes this was with DLSS4 and MFG on the 5070 vs older DLSS+FG on the 4090. And the important part here is MFG which can generate up to three fake frames per each real calculated one. Normal FG can only do one fake frame per calculated one. This means that just with MFG, the new 50 series GPUs can nominally gain three times the amount of fake frames compared to the 40 series. That way Nvidia can probably look at a frame output number and say the 5070 puts out the same number of FPS as the 4090. HOWEVER (!!!) it still means it only delivers 25% of that number in actual playable frames, the rest is just predictive noise. Compared to that, the 4090 offers 50% of playable frames. So in reality the 5070 is about half as powerful as the 4090. Frames generated via MFG or the older FG do NOT add to playability. For all intents and purposes it's just lipstick on a pig to make the game look more fluid by adding frames to it but the games will never run better thn what the GPU can actually calculate in realtime, i.e.: the good old "brute force" (as Nvidia put it so nicely) rasterization performance. Anyone who buys into the 5070 = 4090 marketing is going to be very disappointed. Coincidentally I've got a couple bridges to sell if anyone is interested, I'll even trade for your 4090s.
It's very sad I've seen people in comments trying to say that the fake frames are so great and awesome I have a feeling that people that say those things never actually experienced what the input lag is like, and how big the difference to native really is
@@ADepressedGuy-l1p they are great and awesome when feature is used properly. And they'll improve as time goes on. The problem is bad GPUs and terribly optimized that requires you to abuse these techs and then people wish they never existed because it made developers and GPU makers so lazy and lackluster.
+ you can just use something like Lossless Scaling from Steam (or similar (maybe FSR 4) will be able to run too)). May not be as good but you could do that as well in some Way. So Nvidias Marketing Claim is even more BS. The Tech as is may be cool but you will NOT get an RTX 4090 if you buy an RTX 5070. Dont do it. If you do sell the RTX 4090 to me for 1000 Euro, i will make use of it. Trust me.
I was critical of the latency resulting from a 3x frame generation but it seems like nVidia has handled that problem quite well. What remains to be seen is if the serious image quality degradation from all of that predictive noise can be dealt with appropriately. (Even more hilarious when you're running at a real resolution of 1080p! - I don't think I'd spend £2k just to render 1080p frames )
you are correct about fin curve and bend. the product manager told in a interview that the air moves faster in the edges of the fan and slows down the center you get. so it creates a hum and beat (gamersnexus noctua interview has detail). so those are specially designed to keep airflow speed similar while not hurting performance
I don't think we should consider anything but native rendering. Yes many of us use DLSS, but it's not because we get "better performance". Rather it's because we HAVE to use it on modern titles to get "acceptable performance". With game devs leaning into the crutch of AI upscaling for frames as much as Nvidia is leaning into it instead of improving typical hardware areas, the gamer is losing (a smeary, shimmery, ghosty battle).
The bent fins are essentially just a surface area trick. Since they are closest to the heat source, they wanted those fins to have as much area as possible.
Fin profiles are mind blowing. I bet they did do not make 58 kinds of them, less is possible. Seems obvious that more surface near the heat source is better even with less airflow.
On the bends of the heatsink. Can't it be done if all of the sheets are stacked and stamped once? So that the first sheet would be bent the most, then the second one a little less, and so on and so on?
The different fins can also be made with one pressing of aluminium sheet and cut down. You do not need 58 different stamps for it. Just one containing all the 58 in one press
Problem is we kind of got force DLSS onto us, you can tried disable DLSS TAA in modern game and it look horrible and native TAA is also horrible. So the only option is to use either DLSS or FSR. If possible I would love to not use it but even at 1440p today no upscaling looks so horrible.
100% this. Some games run like crap and you need it for decent performance, other games wouldn't need it but you have to turn it on or run no AA because native TAA is not well implemented, and then there are the games where no AA is not an option because the assets are made with the presumption that you would run TAA or an upscaler and look like crap with it off.
Hey Derbauer, The concave nature of the finstack is exactly as you said, it's due to pressure differences when a fan is operational, the air closer to the pivot point of a fan moves slower than air on the outside edge, so to try and balance airflow movement and reduce turbulence and noise, Nvidia has reduced the finstack thickness in the center, above the fans to compensate. I can't confirm the curving of each individual fin, but it appears to be made to make room for the main PCB, if Nvidia didn't curve each fin after the first, then the restriction would've been too much in that particular portion and heat wouldn't be able to escape the heatsink.
Although I kinda understand why Nvidia pushes AI more than raw performance, it is sad to see because if you're buying the xx90 GPUs, chances are you use your computer for more than just gaming, where DLSS does nothing for you.
Tensor cores are used for more than just dlss lol. Ray reconstruction is used for 3D workloads and many more uses. I can say by same logic that gamers will find that 3x9th gen encoder useless unless they stream or record.
@@sheepmasterrace has nothing to with CES or RTX 50 series. The Biden administration is going to put more Data Centre AI GPU sales restriction to other countries. This will hurt revenue Hence why the drop in stock
Stamping on the cooler fins, there might be just one bigger stamp to cut and bend that unique fins at once, it is probably that way, it would be less expensive than making individual 50 something stamp matrices. The build is amazing, that 4090 prototype from GN also.
@@devgoon That's more reason to correct him. As long as it's a friendly correction, and not just being angry because he used the wrong word. It's an honest mistake.
The bend in the radiator fins is there to control the flow of air. As the area with the most bend is closest to the PCB it will be getting the hottest therefore you want to slow the flow down. As you get further away there is no bend which means more flow can pass through it This is to utilise the radiator more efficiently.
One stamping die can make more than a single part, especially with such a simple part. For example with car body panels, one tool makes two panels, LH and RH.
Honestly seems like more traditional AIB partner cards might be easier for water-cooling than this one. This is clearly “purpose-built” and so would be a bad candidate for disassembly.
Cool, now dont buy it and stop crying online. I just wanted +30% RT performanslce from my 4090 to be satisfied with no raster demands and i was very surprised that even $750 5070ti achieves that level. I am speechless, well done Nvidia, leave haters weeping.
I'll take 240 fps with dlss fake frames that looks 96% as good as native, over 240 fps at a lower native resolution that looks worse than the DLSS fake frames.
I'll take 240 fps with dlss fake frames that looks 96% as good as native, over 240 fps at a lower native resolution that looks worse than the DLSS fake frames.
Still using a Seasonic 1500w prime PSU from 2008 in my daily machine and it is rock solid and goes months between restarts, the 12 year warranty for those is well deserved.
So this mean probably no waterblock for the 5090 FE then. But the FE look like the perfect card for SFF cases then. Very tempted to trade my O11 EVO XL in for a SFF case.
@@ionstorm66 isn't the only difference between AIB and FE the cooler and pre Overclocking? You can change the overclocking yourself and if you want to replace the FE cooler for a waterblock anyways, what does it matter if it i just comes with the block preinstalled?
@user-to7ds6sc3p That would be cool if true but unfortunately the PCB on the AIB cards will look nothing like the FE PCB. Especially now that they are using 3 separate PCBs for this FE design, someone will have to make a waterblock specifically designed for the FE.
@user-to7ds6sc3p there is a FE PCB, then a reference PCB Nvidia provides to AIBs that some cards will use, then tons of custom PCBs that AIBs design. They all take different water blocks. The issue with the way the 5090 is designed will make the mechanical engineering on the water much more time consuming.
@@-Devy- I was watching Vex stream live, and trust me, the counter was actually close to 200, they actually beat AMD because there was a large amount of the word AI splattered in multiple slides
if you drive or fly simulators in VR (Virtual Reality), you'll understand why all this AI "oooh aaah" spiel feels like a kick in the balls, because it's raw power (and rasterization, native res) that matters. Good luck making DLSS and FG work good in VR and not throw up your lunch....
The bends in the heatsinks is more than likely to slow down the air flow to probably catch more heat. This is innovation the way they optimized the pcb is ground breaking!
3:38 27% with RT means max 20% in raster. So 20% more performance, for 25% more money, using 30% bigger die, and using 30% more power. This is not impressive at all from the gaming perspective.
Right. Pay 2000$ for a 575W card to play with 50ms latency. That card only exists for reviewers and rich people. Nobody who has to consider value is going to buy it.
Bending the fins increases surface area, in AC we do this by using pleated filters as apposed to flat increasing the surface area , It measurably works. Its a technology that is out there so not much more in fabrication cost.
A lot of things equal job losses. My bigger issue with AI, is that it does its job worse than the humans. It's not there to make things better, its to make it cheaper.
The bends in the heatsink fins are very interesting, my guess would be to increase the air pressure closer to the heat source with added turbulence while having the lower air pressure from the edges help extract that slower hotter air. Another guess would be that by slowing the air nvidia are allowing a more dense environment for heat transfer. If nvidia wanted to experiment with fluid dynamics, they certainly got the hardware available to do it!
I had seen quite a bit of info on this explained to PCWorld but the heat sink fin design flew over my head. What an amazing cooler - compared to some of the almost 2kg lumps the third parties are producing I can’t wait for a performance, noise and thermal comparison! Cheers
It's possible they are using the same stamp die to make these different fin depths by using a stronger force for a deeper bend, and then it gets lighter and lighter for the shallower stamps, OR a light force that stamped repeatedly to make the bend deeper and less stamping for the shallow ones.
When the title said "reverse-engineered", for a moment I thought Roman had literally built a version of the RTX5090 by analysing the input and output signals from the components... 😂
awesome design , all the heat pipes get to do the same work, short runs to the fins. Only concern would be the 4k signal through a connector on the board and through a ribbon cable.
what's cool about the 3 piece pcb design, is that if the connector breaks due to sag, you can just replace the pcb connector, it also reduces the problem due to sag the gpu die might rip the pads under it, or in the rams too. also in the future the pcb could be just mounted on a motherboard by the connectors it has on the back like a raspberry CM or any compute module really, we mich see SFF pc's using that, or perhaps a thin PC using it
Clearly sag poses less of a threat to the die or MCs with this design, but my first thought was that it's putting all the weight on a much smaller area and a much smaller connector, which I'd think is definitely not a good thing. No one seems to care, though, and nvidia's engineers probably wouldn't make stupid design choices on a flagship product with so much R&D, so what am I missing?
i used to work in a factory that made radiators at a larger scale and was meant for industrial purposes usually. but the types of dimples and how the are placed around pipes and on fins will change air flow dynamic. from what i can see its more likely to help with more air contact space closer to the hotspot of it. its gonna be a hot gpu for sure if they didnt nail it though.
14:29 i think the curve in the middle of the heatsink is to compensate for the missing surface area as a result of curving the top of the fins. the bigger curvature in the middle probably allows for same surface area PER metal fin even though the curvature increases as it goes towards the middle of the card.
Your assumption about air resistance and why the heatsink is concave is correst, saw Nvidia engineer talk to PC World about it. More air pressure comes from the outside of the fans and whatnot.
One thing Nvidia could be using to reduce tooling cost is "progressive stamping" where each hit of the die causes the sheet metal to deform slightly more. Another solution could be an "insert" in the tool, where only the little profile of the bump in the fin is being changed out. Maybe a combination of 5 separate inserts, being used progressively over 10 strokes each? Still more expensive and time consuming than using one and done! The whole flow through design and overall smaller size is very very cool!
I'm buying it just for the cooler design. I'm using the DerBauer 011 case. My GPU right now blocks the airflow from the fans on my bottom radiator through to the fans on my top radiator. This design would allow the air to flow right through. Honestly, it's genius.
your AI mentions per minute chart is hilarious.....love it because I am becoming hyper-sensitive to hearing the term almost like finger nails on a blackboard!!!!!
you don't necessarily need a whole separate stamping tool for each variant, you can have an insert for the bend that can be swapped separately from the whole tool
The FE cooler is honestly the most impressive part of the 5090.
Why?
@@ExtremeGamersBenchmarks 2 slot, sff ready, design. although we'll have to wait until benchmarks come out to really see if its as good as normal air cooled designs
And the tiny pcb
Not the insane 512 bit memory bus that was last used 16 years ago by GTX 285 because it's so difficult to implement? And the fact that they managed to fit it onto a small PCB too...
@@ExtremeGamersBenchmarks lets not forget that its also a 600W, so the fact that its just a 2 slot gpu for such power, the cooling is impressive for its size, hopefully its not a bottleneck
😺 Timestamps
12:17 First meow
12:30 Second Meow
12:40 Exclusive Interview
Thanks
2:15 First Appear
Doing Bastet's holy work.
CRONCH
I am here for plant now but thanks
Thinking like an engineer, you may not need 58 different tools. If all the bends have the same radius, they could just use the same tool, but stamp the different parts a different number of times for each piece. You can calculate the deformation at a given pressure for the fin material, so for example, the least bent fin gets 1 stamp press, and the most gets n+58 stamps. The process becomes longer and more involved, but the cost of the tooling could remain within reason.
Great video, very interesting, I always love your work!
My thoughts exactly. Doesn't make the design any less impressive though.
Thank you. That's exactly what I immediately thought looking at those fins. Nobody in their right mind would make 58 different tools just for a single radiator, instead the shape of those fins look like something you can make with one customizable tool. Which is still pretty cool and much more cost efficient. It might work just as you described or it might have different customizable pressure parameters for each desired deformation of a fin and just cycles through them.
Could they also be stacking up the fin plates on top of each other and then stamping them all at once? The top stamp would be the sharpest, and all others after that would be greater radius as they all mold around each other in the press.
@@billgaudette5524 They will no doubt be re-using the stamping tools in the other FE coolers as well, just not so many fins.
But but but get hyped. I believe this thing will throttle in a small form factor case. Definitely look out for reviews with open benches.
30% more performance with 30% more cuda cores and 30% more power.
👍
And 30% smaller overall dimensions perhaps? 😊
Edit: Had to edit it because people got made the way I used the word package.
@@TheLastTaterthe die size is about 25% larger
Doing the minimum expected improvement per generation only - like Intel used to do with CPUs
@@anone3842 amd struggling to be the only competition in both markets as soon as they started to do well in CPU's they got slammed on the GPU side...
But only 20% increase in MSRP! Progress!
"Excuse me, did you want to say anything in the video?"
*cronch* :3
12:39
14:08 PCWorld had an interview, with an Nvidia rep, confirming the curve's purpose matching your suspicions. Good catch!
I was about to mention, PCWorld's "Nvidia Talks RTX 5090 Founders Edition Design" explains the concave 'grill' shape, the fin variations, and the triple PCBs. And, indeed, Roman pretty well nailed them all. Other items in that video are the purpose of the side vent, which is not just another exhaust point, as well as fan blade design, and possibly a couple other elements I'm forgetting.
@@A_GWAKwork_Orangenoctua money is right up there. 💰💰💰
Only reason I can think of for someone to a) employ an ex EK waterblock designer and b) get them to do an accurate CAD model of 5090... is to make future 5090 waterblocks
The waterblock manufacturers have long had the AIB's board designs from the AIB's with the full knowledge of Nvidia. They are likely to be available on launch day.
Thermal Grizzly GPU waterblocks.... YEP, take my money! 💰
I can think of another reason - because 3D design skills are transferable and there was a sudden oversupply so to speak of ex EK employees looking for work that the tech RUclips community specifically went out of their way to find jobs for
PCWorld has an interview with a NVIDIA product manager yesterday which confirms many of the assertions raised here; the fins are concave to optimize for the pressure delta of the passthrough fans, the angled fins are to redirect air to make a "shield" to deter cannibalistic looping air. Definitely worth a watch.
It makes me very happy to see Joe here and that he's landed on his feet with you.
Tell me Joe is doing all that work because there is a Grizzly FE block in the works , don't waste this man talent . By the way we need an entire WC lineup from Grizzly !
1:32 that "whatever" came from a man who is completely done with A.I marketing.
this is my united states of whatever.
He better get used to it. AI isn't going anywhere, we're entering a new phase of graphics rendering and generation. If we want Hollywood grade movie CGI running in real time on a single PC, we're going to have to rely on AI. We're nowhere close to be able to brute force those level of graphics in real time.
arent we all
@@03chrisvNah, we good.
@@03chrisv Sorry, but modern day graphics can only ever be called brute forcing compared to 8th generation consoles. Ryse Son of Rome anyone? God of war on Ps4? Absolutely gorgeous games that run amazing compared to modern day SLOP.
Makita eating the mic killed me
* CRUNCH *
I bet he has another cat named dewalt or milwaukee
@@bakatenchu Of course, it has to be bosch. It is German.
1:19 graph is all we were waiting for.
yep we are all here like stop the AI bs and hes like: better luck next time
Amd doing a strong showing there. You tell why pat was out
nice benchmark graph amd is winning it seems
Clearly AMD is winning! XD
RTX 5090:
30% faster than 4090 (ignoring framegen, pure raster)
30% more power consumption than 4090
30% more expensive
30% larger die size on the same process node
Such innovation, much wow.
Crazy is the 2080 ti with dedicated Tensor cores vs 1080ti had a higher delta gain. Hey at least we can rely on smoke and mirrors to make up for the marketing and hype shortfalls. :kookoo:
Pure raster performance has been dead for awhile now
@@aberkae the 2080ti had about a 25% delta in raster, that's not higher than 30%. The msrp from 1080ti was $700 and 2080ti was $1000 usd, that's 30%. so we got 25% more performance for 30% more money back then. The 5090 has a higher delta gain to the 4090
@@aberkae TU102 was also 60% larger than GP102
@Stef3m So Nvidia spoiled us until now.
As someone who works in a metalshop with laser and punching machines, what they most likely did to get the decreasing profile in the fins, is use the same tool but use less force while punching in the profile. It's possible they could use some Ertalon (nylon) in the die to have some support when the profile gets smaller. We use wheel forges in CNC punching machines that roll over the plate to make shapes like that (but much longer) and then lower the pressure so it will flare out the profile into the metal sheet and makes it look better.
Usually the dimples, bends or louvers would be to increase turbulence and thus increase the air side heat transfer coefficient. The concavity of the fins may be to try and increase flow velocity through the center by increasing the resistance on the edges (since there is more fin surface area).
Honestly the thermal design of this card looks amazing and a real treat for thermal engineers. I hope I can get my hands on one to actually see it in person.
I wonder if they have brazed the heat pipes to the vapor chamber or they have them directly connected to the chamber. Can't wait for you or Steve to cut them open!
I think dimples serve to decrease turbulence, because why would you want turbulence. I know from practice when i put 4090s on a rig 20cm away from each other they were overheating and when i put my hand between them there was huge turbulence. Air was not going the way it was intended to. It seemed like fans were as much sucking as blowing the air.
It makes me think that if you want to make a PC rig that has exceptional thermals you better do it flawlessly or give up all altogether. When you have weak airflow atleast you didn't create turbulence.
I bet you the primary reason for the dimples is optimizing air flow and reducing pressure when multiple cards are stacked together. Remember, the 5090 is only a gaming card in name at this point. A lot of these will end up getting tossed into AI farms. The partner cards will likely get stripped again like what happened with the 4090. That's why most the cards on ebay are listed as "No core or VRAM". Jensen knows exactly what he's doing.
I think they said something with 3d vapor chamber my guess is like you said
The vapor chamber hugs the heatpipes for more surface area and is probably brazed as well
Since looking at the design they took a WHOLE HECK load of effort into making it a 2 slot card
As Nvidia is becoming more inclined to thinner cards third parties are producing bigger cards than ever
3rd parties have to justify their ever-increasing premiums. The only thing they have left to market to rational customers outside of mild factory OCs are quieter and cooler GPUs
They've got some slim water cooled ones too... but some of these 4 slot cards just look comical.
@@zacharytaylor8523 ?? AIBs have paper thin margins on graphics cards.
@@zacharytaylor8523 they cn jusify their pricea by not beeing sold out
@@auturgicflosculator2183They look really good to me. This FE just looks boring. Also, AIBs improved greatly along with FE for years. I am sure that more premium models like astral, master will be much colder than this FE.
15:30 Nvidia can use only 1 stamping tool, Basically stacking the sheets and stamping with a curvature value. The highest sheet will receive the most curvature while the lowest in the stack will be minimal.
regarding the stamping tool, I doubt that they made 52 different stamping tools, I would say it's more likely that it is stamped with varying pressure with the same tool to get the different bends. Still an expensive process as they need to keep all the plates organized for assembly though, even if they only used one tool.
They could create a tool which stamps all 52 fins in one strike, and then somehow separate individual fins from that single sheet.
@@asmi06 yep
@@asmi06 Was thinking about this too, perhaps there's a manufacturing technique which allows for stacking multiple plates at the same time, reducing the cost of producing parts like this significantly?
They would have needed 58 tools because you also have the feature on the top for the spherical dent where each sheet would have a different radius. Also if you use the same tool with varying pressure your thickness in the bend would change. 58 tools at 5K each is only 290k, it is a drop in the bucket, the real cost would be in assembly, where you would need a machine that automatically stacks all 58 sheets in the right order before pressing the heat pipes in. They already have a couple stamps for the triangle portion so it's no big deal. The cost is also divided across 4 models on the consumer side, and may also carry over to their B series pro cards in some capacity. Looks like NVIDIA spent more time and money designing this cooler than AMD did on their RDNA 4 architecture, or Intel on Arrow Lake.
@@asmi06 I've actually done this in a press shop. Each section of the larger die would have its own specific section designed to make one fin in an assembly like this one. You simply increase the surface area of the entire die and use a larger machine. Individual fins could be cut from the resulting sheet afterwards easily. These fins are quite thin as well, making everything operate on a smaller scale of machinery needed which would lower the cost of tooling investment.
Regarding the shape of each fin being slighty different: Perhaps the method used is to stack the fins and stamp them together in a single step, with a single two-part mold. The added thickness of each fin would then ensure each one is a little bit different. Think of it as if the stamping mold is for doing a "D" shaped curve, and the width/thickness of the "D" itself is given by the added fins.
The keynote's images seem to confirm this method. (maybe some sort of liquid is sprayed between the fins first to allow easy separation afterwards, but that is a detail of the process anyway).
If this does not aling exactly with what we see in the images, maybe there is some kind of sacrificial separator between each fin, or those are the fins that go to the other side of the cooler...
This was my immediate thought too.
Incisive commentary from Makita.
Now that the EK Joe is working with Der8auer there will be waterblocks for 50xx from ThermalGrizzly
Not a bad thing. 🍻
PC World did a video with 'Justin Walker - Senior Director Of Products' where they went over the card and explained about the PCIE board, where it slots in, and also the PCB of the IO and how and where that fits into the card. The two slots at the top are for airflow, it's angled to blow air in a way so that air cannot wash back around and get sucked back into the fans, effectively creating an air barrier. And yes you are correct, the fins are concaved for pressure reasons.
As an engineer that specializes in cooling, the fin designs are likely serving a few purposes. Varying the length of the fins with the concave shape to the overall fin stack allows them to control heat transfer to the fins from the heat pipes in a very efficient way. You would want less heat transferred to the area above the fan hub for example because there is significantly less flow over those fins. This is probably the same reason for the increasingly round indent in the fins toward the center of the card, except in that case they are using the fin shape to slow down the flow and add more surface area for heat transfer to take place. Just my 2 cents. Great video!
Good to see watercool Jesus found new home.
He must meet Tech Jesus
Roman please make this happen
@@iamdmc Steve was there, so they met.
@@iamdmc I believe Tech Jesus connected him to Der8auer/Roman during the EK chaos.
@@iamdmc I've met Steve a few times already, we should be appearing together soon.
LMFAO, “Excuse me did you want to say something?” Cat response *eats mic*
When you do your review, can you show PCIE 4.0 VS PCIE 5.0 I'm on a 5800x3D so would like to know thanks 🙏
I’m on 5800x3D as well. Still a beast of a cpu. I play all my games at 120hz and have never been cpu throttled.
Genuine question: why would you buy a 5090 if you're running a 5000 series AMD processor? 😂
@@deejnutz2068 well, in my case with the 5800x3D and an Rtx 4070 the limiting factor in all games I play is always my graphics card. I play everything in 4K bc I’m a couch gamer. The 4070 does pretty good for most of my games “most games I can do 4K 120hz on medium-high settings” but I’ll likely get a 5080 if I can get one for msrp. Also I never see my cpu usage go above 20-30% usage. So genuinely no reason to upgrade to am5 yet.
Hopefully it will be 1-3% difference like last time when they tested Gen 4 cards on Gen 3 boards
I also downloaded the SVG chart from NVIDIA and broke it down to a pixel perfect count. My result showed a 27.56% uplift over the 4090 in the Far Cry 6 benchmark.
Is that the former EK guy?
yup, thats for sure one of the old EK dudes.. im glad that he is working with Roman. This is good things.
He moved to germany to work with thermal grizzly? That's great!
Thats not the former EK guy. That is THE guy.
@@AttilaGoborlol
@@AttilaGoboryou are so spot on…future looks good!
about the heatsink, all the aluminum sheets are probably pressed against a cylindrical object (probably made of steel) AFTER they're stamped.
next nvidia generation will be named AI 60XX. with this they will win next CES AI competition.
hopeful delirium
Maybe
AI 60xx TIAI?
You double the performance!
@@haukikannel to double the performance, they need to add X somewhere. So 60XX TIAIX maybe.
15:15 the different curve in each fin is possible with a single press. In a stack of fins, the closest to the press will receive the most deflection and it tapers off exactly as you see it on the card, due to each fin absorbing some deflection by compressing.
I can picture what you mean. That sounds plausible.
You don't need different tools for the differently bent fins, just press the tool deeper for each progressively further bent fin. Any CNC stamping tool could do this.
Every now and then, I feel then urge to pop on here and thank you, once again, for the time and work you put in to produce both DE and EN videos. You could have taken shortcuts to accomplish this or chose one language over the other to the dismay of your audience, but you chose the quality path and we appreciate you for it!
Dude, you don't need clickbait thumbnails. Don't hold a card that you don't have.
For the 58 different fin designs, it could be that they designed a wide stamping tool with an extrusion for those curved bits that goes from a larger/deeper extrusion on one side, and slopes up to a shallower extrusion on the other. Load up a wide (fin width x 58) sheet of metal in there, and stamp out ALL 58 fins in one go. Once stamped, the sheet would get machined apart into the 58 separate fins, and that's that.
This would result in fins whose extrusion depth changes across each individual fin's width, but if all fins change similarly, then that shouldn't be a problem.
5:59 frame generation is not more performance. Better frame smoothness, but system latency is even higher which goes counter to the conventional measurement of performance where higher FPS means better responsiveness.
And you also have reflex 2 which is insane and will provide excellent latency regardless of frame generation.
Fake frames are analogous to an automatic transmission if you think about it. Automatic transmission can't anticipate the road, just like how the gamer will change the scene.
It's an instantaneous smoothing.
And all the fastest cars are? Automatic/electronic shifting, and what do they all have? Assistance, ie traction control, differentials, abs, and so on.
Nvidia might be adding latency but I assume their but unveil of reflex 2 is to counter this, much of what they spoke about was the use of ai to render frames in a more accurate light to how it should be.
Comparing a car to a computer is already a stretch and it's quite a pointless analogy, nobody needed an analogy.
And you can't actually state it has higher latency without seeing it run, my guess is DLSS 4 limits the amount of native frames slightly, so rather than struggle to render 40 it does 30 and reduces latency through this and reflex 2, Nvidia may be greedy and may be cornering the market, but they aren't stupid, DLSS, even 1.0 was quite a piece of software, and they rarely mess up, I highly doubt that they won't deliver on this occasion.
Most people moaning about latency and don't even benefit from 5ms, let alone the 20s they so badly dread, if it was poor, only a small portion would truly be affected in any forms meaningful.
@erenyaeger7662They cannot change the fact that 75% of the frames cannot respond to your input. Reflex will do nothing to help that.
@@fewik856730fps response times feel atrocious no matter how slow your brain is.
From what I've seen the different bent radiuses indicate that they don't cut the fins from single sheet but rather cut them in one big stack. This is why the radius changes. Each fin has increased radius by the thickness of the fin.
This will be the rarest model again, together with HOF. Scalpers are already rubbing their hands.
I mean...Nvidia is the second-most valuable corporation in the entire world, I'm pretty sure they can afford a few dozen or even a few thousand stamping tools if they want. That's not even lunch money to them :)
Anyway, I really just wanted to say that I always appreciate the depth and detail of your videos, and the fact that you shoot every video twice. I especially appreciate how you bring your knowledge and experience of the manufacturing and retail aspects of production and products into your videos - these are things that most other content creators can only speculate on, whereas you bring years and years of first-hand experience
THANK YOU!!!
Impressive! I would love to see if you can technically dissipate 575watt of heat with two fans by just heating up a block with the same amount of heat pipes and two 120 fans to do a proof of concept. I know the fins would be impossible to reproduce but that should give us an idea if it’s really possible that nvidia might actually be onto something or they are compromising performance to meet their SFX requirements.
3 fans ...it is a 3 fan design, there is one hidden in the center
I think it's possible and even likely that they made a stamping tool with 57 different shims that can be placed over it in order to get 58 different depths of the curve. stamp 1000 or so at one depth, add a shim, stamp 1000 more at the next depth, and so on.
Regarding the 5070 = 4090 marketing nonsense: according to Nvidia's own footnotes this was with DLSS4 and MFG on the 5070 vs older DLSS+FG on the 4090. And the important part here is MFG which can generate up to three fake frames per each real calculated one. Normal FG can only do one fake frame per calculated one. This means that just with MFG, the new 50 series GPUs can nominally gain three times the amount of fake frames compared to the 40 series. That way Nvidia can probably look at a frame output number and say the 5070 puts out the same number of FPS as the 4090.
HOWEVER (!!!) it still means it only delivers 25% of that number in actual playable frames, the rest is just predictive noise. Compared to that, the 4090 offers 50% of playable frames. So in reality the 5070 is about half as powerful as the 4090. Frames generated via MFG or the older FG do NOT add to playability. For all intents and purposes it's just lipstick on a pig to make the game look more fluid by adding frames to it but the games will never run better thn what the GPU can actually calculate in realtime, i.e.: the good old "brute force" (as Nvidia put it so nicely) rasterization performance.
Anyone who buys into the 5070 = 4090 marketing is going to be very disappointed. Coincidentally I've got a couple bridges to sell if anyone is interested, I'll even trade for your 4090s.
It's very sad
I've seen people in comments trying to say that the fake frames are so great and awesome
I have a feeling that people that say those things never actually experienced what the input lag is like, and how big the difference to native really is
I like to say, "I've got some lovely oceanfront property in Kansas for sale."
@@ADepressedGuy-l1p they are great and awesome when feature is used properly. And they'll improve as time goes on. The problem is bad GPUs and terribly optimized that requires you to abuse these techs and then people wish they never existed because it made developers and GPU makers so lazy and lackluster.
+ you can just use something like Lossless Scaling from Steam (or similar (maybe FSR 4) will be able to run too)). May not be as good but you could do that as well in some Way. So Nvidias Marketing Claim is even more BS.
The Tech as is may be cool but you will NOT get an RTX 4090 if you buy an RTX 5070.
Dont do it.
If you do sell the RTX 4090 to me for 1000 Euro, i will make use of it. Trust me.
I was critical of the latency resulting from a 3x frame generation but it seems like nVidia has handled that problem quite well. What remains to be seen is if the serious image quality degradation from all of that predictive noise can be dealt with appropriately. (Even more hilarious when you're running at a real resolution of 1080p! - I don't think I'd spend £2k just to render 1080p frames )
you are correct about fin curve and bend. the product manager told in a interview that the air moves faster in the edges of the fan and slows down the center you get. so it creates a hum and beat (gamersnexus noctua interview has detail). so those are specially designed to keep airflow speed similar while not hurting performance
Running different scenarios for a comparison graph is really dirty.
This is the only video I was waiting for after the announcement of the RTX 50 series !
I don't think we should consider anything but native rendering. Yes many of us use DLSS, but it's not because we get "better performance". Rather it's because we HAVE to use it on modern titles to get "acceptable performance". With game devs leaning into the crutch of AI upscaling for frames as much as Nvidia is leaning into it instead of improving typical hardware areas, the gamer is losing (a smeary, shimmery, ghosty battle).
Exactly! And game publishers wonder why most hours of gaming are done in older games............. 🤔
The bent fins are essentially just a surface area trick. Since they are closest to the heat source, they wanted those fins to have as much area as possible.
16:21 $500k for nvidia is like lunch money.
More like finding a few cents on the ground considering nvidias net worth😂
Fin profiles are mind blowing. I bet they did do not make 58 kinds of them, less is possible.
Seems obvious that more surface near the heat source is better even with less airflow.
On the bends of the heatsink. Can't it be done if all of the sheets are stacked and stamped once? So that the first sheet would be bent the most, then the second one a little less, and so on and so on?
Or have a bunch of bends on one stamping, then another stamping to cut them into individual fins?
The different fins can also be made with one pressing of aluminium sheet and cut down. You do not need 58 different stamps for it. Just one containing all the 58 in one press
Problem is we kind of got force DLSS onto us, you can tried disable DLSS TAA in modern game and it look horrible and native TAA is also horrible. So the only option is to use either DLSS or FSR. If possible I would love to not use it but even at 1440p today no upscaling looks so horrible.
100% this. Some games run like crap and you need it for decent performance, other games wouldn't need it but you have to turn it on or run no AA because native TAA is not well implemented, and then there are the games where no AA is not an option because the assets are made with the presumption that you would run TAA or an upscaler and look like crap with it off.
@@DinoTrollerino alternative is a 300GB+ game with high quality assets I guess
Hey Derbauer,
The concave nature of the finstack is exactly as you said, it's due to pressure differences when a fan is operational, the air closer to the pivot point of a fan moves slower than air on the outside edge, so to try and balance airflow movement and reduce turbulence and noise, Nvidia has reduced the finstack thickness in the center, above the fans to compensate.
I can't confirm the curving of each individual fin, but it appears to be made to make room for the main PCB, if Nvidia didn't curve each fin after the first, then the restriction would've been too much in that particular portion and heat wouldn't be able to escape the heatsink.
Although I kinda understand why Nvidia pushes AI more than raw performance, it is sad to see because if you're buying the xx90 GPUs, chances are you use your computer for more than just gaming, where DLSS does nothing for you.
That’s probably why Nvidia stock suffered
Imo, those that need the extra work performance will just go for the Nvidia Quadro line
Tensor cores are used for more than just dlss lol. Ray reconstruction is used for 3D workloads and many more uses. I can say by same logic that gamers will find that 3x9th gen encoder useless unless they stream or record.
Incorrect.
Dlss runs on Tensor Cores.
Tensor Cores have metric ton of uses
@@sheepmasterrace has nothing to with CES or RTX 50 series.
The Biden administration is going to put more Data Centre AI GPU sales restriction to other countries.
This will hurt revenue
Hence why the drop in stock
Stamping on the cooler fins, there might be just one bigger stamp to cut and bend that unique fins at once, it is probably that way, it would be less expensive than making individual 50 something stamp matrices. The build is amazing, that 4090 prototype from GN also.
Reverse engineered is probably the wrong wording to use. That implies you built your own physical 5090
yep 100% this.
Yea, "dismantle" or "deconstruct" would be a better description.
English is not his first language. Let's give him some slack.
@@devgoon That's more reason to correct him. As long as it's a friendly correction, and not just being angry because he used the wrong word. It's an honest mistake.
Analyze!
The bend in the radiator fins is there to control the flow of air. As the area with the most bend is closest to the PCB it will be getting the hottest therefore you want to slow the flow down. As you get further away there is no bend which means more flow can pass through it
This is to utilise the radiator more efficiently.
the engineer team that did this are masterclass omg. makes me actually consider a FE
The small bends in the fins is for more surface area, probably compensating for the 2 large sunk dimples over in the heat sinks over the fans.
The word AI being spoken 140 times in 91 minutes sums up big tech in 2025. Like a bunch of mindless squawking seagulls at the beach.
One stamping die can make more than a single part, especially with such a simple part. For example with car body panels, one tool makes two panels, LH and RH.
I'm worried about how they will make water blocks for this design.
Honestly seems like more traditional AIB partner cards might be easier for water-cooling than this one. This is clearly “purpose-built” and so would be a bad candidate for disassembly.
@@stillmoms Yeah, I didn't think of that. I'll be buying mine from Asus. I feel better now!
They won't. AIB's already have blocks available with 5090's pre-installed
@@ATestamentToTech Perhaps Optimus will make one. I hate AIO stuff and only have open loops in my PCs.
@@molly_mallard They are open loop blocks! Pre-installed. Both MSI and Gigabyte already have them
This was such a fun video! So many tongue-in-cheek moments, but also legit info.
@5:40 Fake frames ARE fake frames...
Cool, now dont buy it and stop crying online. I just wanted +30% RT performanslce from my 4090 to be satisfied with no raster demands and i was very surprised that even $750 5070ti achieves that level. I am speechless, well done Nvidia, leave haters weeping.
@@filippetrovic845 Have you been drinking again? Do you even know what I'm saying?!
@@detmer87just another consumer who consumes
I'll take 240 fps with dlss fake frames that looks 96% as good as native, over 240 fps at a lower native resolution that looks worse than the DLSS fake frames.
I'll take 240 fps with dlss fake frames that looks 96% as good as native, over 240 fps at a lower native resolution that looks worse than the DLSS fake frames.
Still using a Seasonic 1500w prime PSU from 2008 in my daily machine and it is rock solid and goes months between restarts, the 12 year warranty for those is well deserved.
If you're using DLSS and the majority of frames are fake...how can you assess the raw performance?😬
Who need raw performance when AI can create as many frames as you like!
😂
@@yamuiemata you can’t. We can no longer use frames as a direct metric for performance.
@@haukikannelbut latency is bad. On fake frames
@@povgameswe can however use compute tasks. Then increase is just under 20% with lower end cards being even lower improvement.
Thanks for noticing and analyzing the cooling fins. And for the meows!
So this mean probably no waterblock for the 5090 FE then. But the FE look like the perfect card for SFF cases then. Very tempted to trade my O11 EVO XL in for a SFF case.
@@irispettson those are AIB cards not FE.
FE blocks should still be possible, but some considerations will have to be taken. Who knows if FE supply will be good or not this generation though.
@@ionstorm66 isn't the only difference between AIB and FE the cooler and pre Overclocking?
You can change the overclocking yourself and if you want to replace the FE cooler for a waterblock anyways, what does it matter if it i just comes with the block preinstalled?
@user-to7ds6sc3p That would be cool if true but unfortunately the PCB on the AIB cards will look nothing like the FE PCB. Especially now that they are using 3 separate PCBs for this FE design, someone will have to make a waterblock specifically designed for the FE.
@user-to7ds6sc3p there is a FE PCB, then a reference PCB Nvidia provides to AIBs that some cards will use, then tons of custom PCBs that AIBs design. They all take different water blocks.
The issue with the way the 5090 is designed will make the mechanical engineering on the water much more time consuming.
One bonus for the separate PCIe board is that there's less chance to damage the GPU solder joints from GPU sag
The AI counter would be higher because you should also include the slides/background images with the word AI too. It was like 220 something lol
Pretty sure he said previously that those were included.
@@-Devy- I was watching Vex stream live, and trust me, the counter was actually close to 200, they actually beat AMD because there was a large amount of the word AI splattered in multiple slides
if you drive or fly simulators in VR (Virtual Reality), you'll understand why all this AI "oooh aaah" spiel feels like a kick in the balls, because it's raw power (and rasterization, native res) that matters. Good luck making DLSS and FG work good in VR and not throw up your lunch....
The bends in the heatsinks is more than likely to slow down the air flow to probably catch more heat.
This is innovation the way they optimized the pcb is ground breaking!
3:38 27% with RT means max 20% in raster.
So 20% more performance, for 25% more money, using 30% bigger die, and using 30% more power.
This is not impressive at all from the gaming perspective.
12:39 best "in the field" interview ever!
Right. Pay 2000$ for a 575W card to play with 50ms latency.
That card only exists for reviewers and rich people. Nobody who has to consider value is going to buy it.
Bending the fins increases surface area, in AC we do this by using pleated filters as apposed to flat increasing the surface area , It measurably works. Its a technology that is out there so not much more in fabrication cost.
A.I = job losses
Hunt or be hunted. Rules of nature.
@@LiterallyLozylTo be, or not to be, that is the question
Wouldn't be amazing if they could use Ai to stop scalpers from buying Pallets of GPUs holding the m hostage.
A lot of things equal job losses.
My bigger issue with AI, is that it does its job worse than the humans. It's not there to make things better, its to make it cheaper.
@@LiterallyLozylWhat or who to hunt if a human cannot compete against multibillion company?
Me when video card shopping: _SAY A.I. AGAIN I DARE YOU I DOUBLE DOG DARE MF. SAY A.I. ONE MORE GD TIME_
Yoooo
The bends in the heatsink fins are very interesting, my guess would be to increase the air pressure closer to the heat source with added turbulence while having the lower air pressure from the edges help extract that slower hotter air. Another guess would be that by slowing the air nvidia are allowing a more dense environment for heat transfer.
If nvidia wanted to experiment with fluid dynamics, they certainly got the hardware available to do it!
This design choice once again lives to show that when the PCB is not obstructing airflow, the cooling performance can be effectively increased.
I had seen quite a bit of info on this explained to PCWorld but the heat sink fin design flew over my head. What an amazing cooler - compared to some of the almost 2kg lumps the third parties are producing I can’t wait for a performance, noise and thermal comparison! Cheers
Im happy with 4090. Will be waiting for 6090
It's possible they are using the same stamp die to make these different fin depths by using a stronger force for a deeper bend, and then it gets lighter and lighter for the shallower stamps, OR a light force that stamped repeatedly to make the bend deeper and less stamping for the shallow ones.
When the title said "reverse-engineered", for a moment I thought Roman had literally built a version of the RTX5090 by analysing the input and output signals from the components... 😂
awesome design , all the heat pipes get to do the same work, short runs to the fins. Only concern would be the 4k signal through a connector on the board and through a ribbon cable.
what's cool about the 3 piece pcb design, is that if the connector breaks due to sag, you can just replace the pcb connector, it also reduces the problem due to sag the gpu die might rip the pads under it, or in the rams too. also in the future the pcb could be just mounted on a motherboard by the connectors it has on the back like a raspberry CM or any compute module really, we mich see SFF pc's using that, or perhaps a thin PC using it
Clearly sag poses less of a threat to the die or MCs with this design, but my first thought was that it's putting all the weight on a much smaller area and a much smaller connector, which I'd think is definitely not a good thing. No one seems to care, though, and nvidia's engineers probably wouldn't make stupid design choices on a flagship product with so much R&D, so what am I missing?
i used to work in a factory that made radiators at a larger scale and was meant for industrial purposes usually. but the types of dimples and how the are placed around pipes and on fins will change air flow dynamic. from what i can see its more likely to help with more air contact space closer to the hotspot of it. its gonna be a hot gpu for sure if they didnt nail it though.
14:29 i think the curve in the middle of the heatsink is to compensate for the missing surface area as a result of curving the top of the fins. the bigger curvature in the middle probably allows for same surface area PER metal fin even though the curvature increases as it goes towards the middle of the card.
🤣I died when I saw the slide comparing AI mention frequency, beautiful
Great job! I had wondered how that cooler was going to work.
Your assumption about air resistance and why the heatsink is concave is correst, saw Nvidia engineer talk to PC World about it. More air pressure comes from the outside of the fans and whatnot.
Aahahah I spit out my coffee when makita swaloowed the mic 😂😂
One thing Nvidia could be using to reduce tooling cost is "progressive stamping" where each hit of the die causes the sheet metal to deform slightly more. Another solution could be an "insert" in the tool, where only the little profile of the bump in the fin is being changed out. Maybe a combination of 5 separate inserts, being used progressively over 10 strokes each? Still more expensive and time consuming than using one and done! The whole flow through design and overall smaller size is very very cool!
Thanks for the Ai//min chart and Joe's hard work on the renders.
I'm buying it just for the cooler design. I'm using the DerBauer 011 case. My GPU right now blocks the airflow from the fans on my bottom radiator through to the fans on my top radiator. This design would allow the air to flow right through. Honestly, it's genius.
This was the ultimate prediction video. I can't wait to see how close to reality you got.
your AI mentions per minute chart is hilarious.....love it because I am becoming hyper-sensitive to hearing the term almost like finger nails on a blackboard!!!!!
you don't necessarily need a whole separate stamping tool for each variant, you can have an insert for the bend that can be swapped separately from the whole tool