The RTX 5090 and RTX 5080 are receiving their final updates. According to two highly reliable leakers, the RTX 5090 is officially a 575W TDP model, confirming that the new SKU requires significantly more power than its predecessor, the RTX 4090 with TDP of 450W.
According to Kopite, there has also been an update to the RTX 5080 specifications. While the card was long rumored to have a 400W TDP, the final figure is now set at 360W. This change is likely because NVIDIA has confirmed the TDP, as opposed to earlier TGP figures that are higher and represent the maximum power limit required by NVIDIA’s specifications for board partners.
↫ WhyCry at VideoCardz.com
These kinds of batshit insane GPU power power requirements are eventually going to run into the limits of the kind of airflow an ATX case can provide. We’re still putting the airflow stream of GPUs (bottom to top) perpendicular to the airflow through the case (front to back) like it’s 1987, and you’d think at least someone would be thinking about addressing this – especially when a GPU is casually dumping this much heat into the constrained space within a computer case.
I don’t want more glass and gamer lights. I want case makers to hire at least one proper fluid dynamics engineer.
I feel these GPU are a raspberry pi away from being fully enclosed and dedicated gaming systems.
This level of power draw is getting silly though. You could legitimately buy a new midrange PSU to Just power this component
Adurbe,
It does seem silly, but I think it’s justifiable if that component ends up doing most of the work. GPUs are the future of computing because their design lends themselves to scalability better than CPUs can.
In blender, for example, I lament the features like physics simulation that aren’t GPU accelerated. The GPU has so much parallel processing power. The CPU’s cores, while being significantly faster than GPU cores on a per-core basis, are extremely inefficient both in terms of energy and die space that they are nowhere close to the GPU’s parallel computing performance. Obvious we’re very dependent on specific software in order to use GPUs, but capability wise I don’t think it’s a contest, the future of high performance computing is the GPU and not CPUs with lots of cores.
IMHO that doesn’t seem so far fetched. I gather that the newest RPIs have decent performance and it’s conceivable for future games to offload more or even all calculations to the GPU.
That’s not true, since GPU is essentially a SIMD system the computational efficiency comes from the fact that the single instruction decode yields more (FL)OPs. But only specific algorithms actually can utilise this — shading is more-less a best case, something like web server would suck terribly. Moreover, SIMD pressures caches and memory bandwidth very hard since you need a lot of local state, and quite often this is the actual bottleneck. And data buses are very resource hungry, which probably causes an ever growing GPU appetite for power — this also locks their evolution with this of CPUs, which face the same challenges.
So, there are many things GPUs won’t ever do better than CPUs, their scalability is constrained by memory as with everything computational nowadays, and they actually face being replaced by dedicated accelerators in the areas they excel.
Exactly. Apple’s main trick to make their ARM based SoCs so efficient is offloading specialized tasks to power sipping specialized units. Now we also have NPUs in everything. I suspect we will see a lot more of this kind of thing going forward.
It kind of reminds me of the (very) old days when there used to be real performance bottlenecks from essentially everything, and we used to add PCI cards for audio, networking and all that. It wasn’t efficiency driven back then, but it performance driven (decoding a single mp3 on my Pentium 166 MMX would eat up almost all the available capacity in that system, heh), but the result was similar. Add more specialized hardware for specialized tasks. Then we started moving features on to the south and north bridges, and now everything is integrated on die. Still, it’s nice to feel like we are in familiar territory.
mbq,
It is true. Although I admit you can add more GPU-like units to a CPU, which would obviously scale more like a GPU. What makes CPU cores inefficient is the complex superscalar pipeline that needs a lot of die space and energy to do so. This is different than GPU designs that dedicate more transistors to ALU operations using relatively basic cores that are individually slow but can do a whole lot more work efficiently in parallel.
Yes I agree. GPU is best for embarrassingly parallel tasks. It’s just the majority of computational work that games need to do is of this variety. This is what Adurbe meant by “these GPU are a raspberry pi away from being fully enclosed and dedicated gaming systems.” A regular CPU would still handle things like usb input, networking code, and maybe sound. But the vast majority of the heavy lifting in a game including the physics traditionally done on CPUs can be done by the GPU.
GPUs aren’t good at running CPU code, but I think we’re going to see more software get developed to exploit the parallel power of GPUs. You can buy GPUs with loads of memory, it’s going to cost a pretty penny though. GPUs can and do use “unified memory” transparently, however the problem is that the PCI bus itself is a bottleneck. At one time nvidia addressed PCI bottlenecks with a proprietary bus called nvlink but now they prefer to sell larger GPUs with more memory instead.
As CaptainN- points out though this isn’t new. We’ve had hardware accelerators for a very long time, and while these can perform better than programmable cores by hardcoding algorithms in silicon, it’s also one of their biggest drawbacks…And these are very expensive to implement and they cannot be changed in the field. So while in theory you might think they’d replace more generic compute units, in practice they’re only good for very specific tasks that never need updating. I’ll provide examples…
Dedicated hardware acceleration works: bitcoin cryptocurrency miners where the exact same math is repeated billions of times per second over decades…CPUs lost to GPUs, which lost to ASIC implementations. When it comes to bitcoin mining, being fast is so important that ASICs are the only competitive option at this point.
Hardware doesn’t work: full TCP offloading. It technically works but CPUs have gotten so fast relative to memory speeds and CPUs have so many cores that linux’s software implementation can outperform the dedicated hardware implementation. Ironically there can be more overhead in sending the ASIC instructions than just doing the work outright. There’s also the matter of memory protection and synchronization. And furthermore hardware implementations severely limit the use of TCP extensions and updates, which can result in negative repercussions for users and applications that cannot be fixed without buying new hardware. For these reasons, sometimes admins find the software implementation works better than the silicon implementation.
Apple M# CPUs have loads of accelerators to improve performance over the generic compute units, which is beneficial for applications using them. But this model really falls short for things like gaming where developers need generic CPUs and GPUs that can be cheaply reprogrammed by them for any conceivable purpose the game may have. And don’t forget that GPUs have accelerators of their own in the form of RT cores, tensor cores, etc.
So I don’t think GPUs are at risk of being replaced by hard coded accelerators on CPUs or in peripherals.. Now it would be very interesting if you through FPGAs into the mix. If intel or AMD threw an FPGA into all their CPUs and open sourced it…I think that could be a significant game changer. FPGA versus GPU would be an interesting showdown.
Thom Holwerda,
Here here. Sometimes it’s hard to buy non-RGB variants of some of these components We end up forced to pay a premium for “features” we didn’t even want.
I’d like to more case go to basics: simple front to back airflow and less churn inside the case. This is the way rack mount hardware is designed. Rather than having Internal fans haphazardly exchange and circulate air inside the case, it would be better to pass the cool outside air directly over exchange fins with passage ways to limit re-circulation inside the case.
We also need much better standards to control the fans such that the peripherals that need better cooling can ask for it without having to install proprietary software. This is a serious problem today as there are notorious problems with nvidia GPUs and the inability to monitor board temperatures that get critically hot. Personally I deal with it by configuring an aggressive fan profile but it’s really annoying that the parts that generate heat don’t interoperate well with fan controllers.
It would not be hard to develop a suitable hardware standard, but the problem is each manufacturer wants to do their own proprietary thing so they don’t release anything as FOSS and nothing talks together.
I picked my gaming case (Fractal Design Ridge) specifically because it ditches the traditional ATX design and places the GPU in the top half of the case, with its own air flow and two 140mm case fans pulling cold air across the face of the GPU. It’s effectively forcing cold air into the existing GPU fans, giving them more air to move around, while also having ventilation at the top for convection cooling. The GPU’s fans rarely spin up while gaming because the case fans move just as much air as them while being much quieter.
The motherboard is in the lower half of the case, separated from the PSU, and each of those sections has its own ventilation. It’s a great thermal design and makes for a PC roughly the size and shape of the PS5. The only drawback is that it takes several steps to open the case and access everything, but that’s a tradeoff I was willing to make for what is essentially a Steam console anyway.
Yeah, no thanks. I may have air conditioning now, but it’s still work to get heat out of this air cul-de-sac of a room in the summer when my RTX 3060 bandwidth-bottlenecks at ~125W of its 170W maximum in a PC specifically built with “what’s the fastest Ryzen I can afford at 65W TDP?” being the first concern.
I was content with a GTX 750 for a decade. I can wait.
I’m in the same boat, currently with an RTX 3050 6Gb after a 1030GT and GTX 750. The 3050 6Gb part is a 70W TDP, but I was monitoring new products for a couple years to see any device under 100W, where power/heat were the main criteria. Clearly we’re a tiny part of the market both in volume and margin though, not a market that’s well served. The 3050 is running all the games I want just fine though.
On the plus side, you can turn down your central heating a bit (at least if it’s winter).
I do wonder how far away we are from total energy management in a house. In the same way wall sockets and power strips are now provided with USB plugs in them, in the future I’d expect that the cooling requirements of individual component outstrips fans altogether. We’ll plumb GPUs into our home heating, so the heat goes out through the radiators of our home.
Either that, or use thermal energy storage like this: https://blog.sintef.com/energy/thermochemical-energy-storage-the-next-generation-thermal-batteries/
Yeah, it’s science fiction today…but give it 10-20 years.
kcorey,
Agreed. If the heat byproducts are used to heat an environment that otherwise requires heating anyway, then the GPU’s heat output is actually being productive. However I suspect that for most of us that’s not the case for most of the year. There is another place to use that heat: a water heater. A greater percentage of the population needs to heat water year round which makes water is a good place to dump excess GPU heat, but a larger temperature differential is needed and would therefor require active heat pumps. Nevertheless GPU heat output can be used to heat water with a heat pump….but now we are talking about a lot more complexity and expense, not to mention the impracticality for normal users to go the DIY route.
It would be interesting for a company to come out and sell a mass market product to do this. A novel and quiet approach to cooling computers with the heat produced being genuinely beneficial for things like water heating. But I doubt society is ready for this sort of thing. Without scales of economy, nobody’s going to want to buy a specialized computer that hooks up to specialized heaters even with energy gains unless money is no object to them.
Heat is an efficient way to store energy at scale without relying on rare earth materials like batteries. This could offer a solution to the unreliability of other green energy sources…Alas the massive salt heating solar plants like those on the outskirts of nevada have so far suffered from huge budget overruns and failed to be profitable with high expenses compared to PV solar plants that have no moving parts. Still though heat storage plants do have one big advantage over PV’s, their energy is available at any time including night. Because of this heat based grid scale storage may still prove useful in the future.
My biggest gripe about ATX is that it supports the GPU only on one side (the back), and with modern GPUs being three-slots wide, this easily leads to GPU sag bad enough to crack PCBs. If case manufacturers could come up with a form factor that supports the GPU on both sides (if Sun could do it with the XVR-4000 in the Sun Fire V880z, it can be done), it could solve a lot of problems.
About ventilation, the CPU also needs to get air, and cases have indeed grown larger to support a higher airflow. Even Alienware has ditched the obsession with tiny cases and went with an appropriately-sized case for their desktops.
kurkosdr,
+1,000,000
This is such a serious shortcoming and it’s really disappointing that our hardware standards are in this state of affairs.
I actually think we could get away with much smaller cases if we didn’t have to treat the case’s air as a large mixing chamber.
Putting aside water cooling, right now we use fans to mix cool outside air with warm inside air and then more fans on the CPU/GPU/MB to dissipate their heat into the case. Sometimes an exhaust fan pulls hot air of of the case. Although this works and larger cases can make this work better, it seems very inefficient. Why not do away with mixing chambers entirely and redesign heatsinks to be cooled by a continuous streamlined outside airflow being passed directly over them and then have the heated air exit the case directly. By doing away with the inside of the case being a mixing chamber, the case can get a lot smaller. The room itself becomes the mixing chamber, not the computer case.
The ATX specification has had support for both ends of the card for ages.
Back in the 80s/90s with ancient AT and microchannel systems you would see the rails on the other end of the case to guide the board and support at both ends.
It’s mostly the fault of the case manufacturers that ditched the rails/support brackets altogether for ages now.
Xanady Asem,
That predates computers I’ve use. I wonder how well it would hold up to modern GPUs. Even bulky cards of the 80s probably weren’t as hefty as modern GPUs. It could make more sense to re-engineer it with a more forward looking standard. Either way most cards would need to be re-engineered for bracketing. Might as well do something to optimize peripheral airflow at the same time. Two birds with one stone.
Attempts have been made to replace ATX over the years, such as Intel’s BTX and, for PSUs, ATX12VO. They generally don’t catch on and get shunned as irritating “proprietary crap” when they show up in the prebuilt market that Intel tends to court first.
Any solution to replace ATX would likely have to be completely backwards compatible. Network effects are very powerful in this situation.
(And I understand why. One of the first things I do when I get a retrocomputer is make plans to replace as many parts as possible with adapters to high-quality examples of SATA, ATX PSU, etc. parts and, in large part, that’s to maximize repairability if one of the newer, less aged parts does fail.)
Hell, the PC I built this January is using my brother’s old Antec case from around 2010 because I see fewer than two 5.25″ bays as “proprietary crap” too.
ssokolow (Hey, OSNews, U2F/WebAuthn is broken on Firefox!),
I agree, consumers often prefer a “good enough” standard over a better one. However I would argue today’s standards are not good enough for modern GPUs.. They break way too easily and shipping a GPU installed with no foam is a death sentence for it…
https://www.youtube.com/watch?v=6H9XeH8G_mM
I see your point, but in this case you’re probably not going to find a single GPU that fits those 1980 ATX mounting points anyway So in this particular case backwards compatibility with those long form factors probably isn’t very relevant.
A forward looking standard could probably just use the same PCI slots but focus on the physical support that modern cards desperately need.
David Dent patented the ATX standard in 1995. Not the 80s. Perhaps you are refering to AT as in IBM AT.
NaGERST,
I’m trying to find some evidence in spec for peripheral cards being physically supported on all corners but I’m not having much luck. I see some pictures of cards in the AT spec here on page 522, but nothing clear and explicit.
https://archive.org/details/bitsavers_ibmpcat150ferenceMar84_26847525/page/n523/mode/2up
I see plenty of historical info on the motherboards themselves, but not the peripherals.
I’m just curious. I’ve never seen it personally on an ATX and I’m trying to find a picture of what that might have looked like (AT or ATX). I’m getting the impression that fully supported cards wasn’t commonplace. If anyone can find a picture, please send link.
Alfman, while it doesn’t cite a source, the Winn L. Rosch Hardware Bible, Fifth Edition, says that a full-length ISA card is 340.7 mm long and explicitly says that length is defined by the length “from the back edge of the mounting bracket to the far edge mating with the card guide” (that array of slots which physically supports the far end of the card).
It then goes on to say that full-length PCI is specified to be 312mm long and that PCI also defines a “short board” that’s 174.63mm long.
…so it sounds like the near universal use of either “short board” PCI form factors or even smaller cards lulled the case manufacturers into cutting costs by omitting card guides in the post-AT era.
As for “A forward looking standard could probably just use the same PCI slots but focus on the physical support that modern cards desperately need.”, that’s basically what I meant.
Design cases that are standard ATX in any way that affects backwards compatibility, but add some standardized means of reintroducing the “card guides” that were ubiquitous in the era of “components are so big that even extra RAM needs to be on a full-length ISA card and forget about stuff like serial or parallel ports being onboard”.
It’s not as if the modern Noctua tower cooler on my 2023 Ryzen in my Antec case from around 2008 will fit into the 2012 HP prebuilt that I repurposed into a “game console but not a console” and those cases are both members of the ATX family.
Ok nice.
I checked the first four 4090 RTX cards in a search to get a random sampling of lengths.
358 mm ASUS
331 mm Gigabyte
304 mm FE
322 mm MSI
The lengths vary by several centimeters although they could probably have been engineered to something specific like 340.7 mm without issue. Of course these are the big cards, but lower end consumer cards aren’t usually that long. It could be better to standardized height so that a solid rail could hold the top of the card,
Given all the problems that have hit the press with these expansion cards over the past several years…
– Bad airflow
– Trouble fitting non-standard GPUs into non-standard cases.
– Unsupported weight
– Stress on PCIE slot,
– Power connectors with dangerously poor connections
… it could be interesting for the industry to develop a PCI backplane plane that works like PCMCIA for high end GPUs. The hot swappable hard disk caddies and power supplies that servers have are awesome. They are robust, easy to load/unload, securely lock into place, and you don’t even have to open the computer to install the hardware.
Now that I’m thinking about it, it’s probably too wasteful and expensive for consumer grade hardware, to say nothing of the RGB fans who demand form over function. I know there’s no real chance of adoption, but I do like the idea of something robust and foolproof. It seems like $1k+ cards deserve to be backed with stronger engineering than what we have now.
Funny thing is, the hand-me-down Alienware Area 51 (the weird triangle thing) that I have hooked up to the big-screen TV as another “game console but not a console” has a solution for varying card length.
They built a little metal piece that goes onto the end of the 9xx or 10xx-series GeForce they put in to make it reach the bracket which substitutes for an ISA card guide.
(I know because I had to change the one part that requires you to remove the GPU and card support rack… I installed a Noctua to replace the lower intake fan. Aside from that, it’s a joy to work on and has an excellent service manual, so, clearly, Dell came up with a silly over-the-top exterior and then handed it to the same case-design team they use for parts that have to appeal to purchasing managers and IT departments.)
I could easily see a new “ATX+S” (S for support) family of case standards where the case has open space long enough for any existing card with a support rack off the end and then GPU manufacturers are expected to bundle that kind of optional metal bracket to extend the card’s length to interface with it, similar to how PCI and PCIe cards with half-height PCBs tend to include a half-height expansion slot bracket that the end-user can swap on.
It wouldn’t solve the other problems, but it would be a gentle way forward on the mechanical support side of things and should be something that could be built on for future improvements.
but but but….. raytracing, ssao and motion blur? Yeah focus on that instead of raster performance and AMD will take over the gaming crown once again.
Not that it matters, nvidia only want those sweet AI money nowadays