There sorta is a cap in place given that most residential circuits in the US are 15 or 20 amps. We're already not far off of that being a serious concern even on non-crazy builds. Edit: ok maybe crazy builds but still it's not that far fetched on current trajectories.
I don't know if I'd call 1800W (residential circuit limit) builds "non-crazy". Even half that power is rare to see since that's pretty much two top of the line consumer graphics cards paired to a top of the line enterprise CPU.
Two 3090s plus a 64 core Threadripper pro will get you beyond 1000 watts alone. Don’t forget power for the 512gb of ram (200 watts), the motherboard, the fans and hard drives and accessories and gamer lights (well, gamer lights aren’t necessary). That’s definitely a crazy build though.
> Two 3090s plus a 64 core Threadripper pro will get you beyond 1000 watts alone. Don’t forget power for the 512gb of ram (200 watts), the motherboard, the fans and hard drives and accessories
Yes, PEAK instantaneous wattage could easily be well over 1kW for a crazy rig like that, but unless you are specifically torture testing the rig over a long duration, it simply isn’t accurate that you are drawing anywhere near the max continuously under gaming conditions.
> Two 3090s plus a 64 core Threadripper pro will get you beyond 1000 watts alone
3090s are 320/350W each and 64 core Threadripper Pro has a TDP of 280W, just as an FYI.
I said near, we still have some breathing room. But as noted elsewhere, 15 A is a max for short periods, the practical limit for continuous use is more like 1200-1400 W which definitely seems like it might be plausible soon.
Plus other things on the circuit. Fancy ultra bright HDR monitors use a decent chunk of power, and anything else.
> But as noted elsewhere, 15 A is a max for short periods, the practical limit for continuous use is more like 1200-1400 W which definitely seems like it might be plausible soon.
Per NEC in the US, the wire used in a circuit has to be rated to handle a current draw of up to 80% continuously at the specified temperature max for the type of installation.
80% of 15 is 12, the average home in the US has split phase incoming power with a nominal voltage of 120V (1440W) with many as high as 125V (1500W). Note that is current draw over long periods of time, which a gaming rig under heavy gaming won’t usually do. Maybe a mining rig back in the day would, but otherwise it’s just not true you’re continuously drawing anywhere near 1,000+ watts on anything but the most crazy setups.
> I said near, we still have some breathing room.
No, you made a wild ass conjecture not based in fact and when gently questioned on it, you shifted goal posts and made up more conjecture.
1200 W, even for a short period of time, is crazy for a desktop computer.
I have an i9-9900K and an RTX 3080. My computer is also plugged into a Kill-a-watt to measure its power usage.
I ran Prime95's CPU torture test using the max power option while also crypto-mining on the GPU. I peaked at 650 watts for a couple seconds until the CPU began thermal throttling.
If I can't hit 700 watts while trying to use as much energy as possible, I can't imagine what monster system would even touch 1000 watts, assuming we're still talking consumer-grade.
Counting my monitors and the small network switch on my desk, I can hit around 750-800 watts sustained with a 5950X and a 3090 running compute intensive workloads as measured at the wall by a power meter.
So yeah, there's plenty of head space but I have also hit circuit breaker issues recently on a proper (non consumer) workstation so the limit is fresh in my mind.
Don't forget the PSU isn't perfectly efficient; it's usually between 80 and 90 percent. So for safety, about 1100W (for the total of the DC power draws of the components fed by the PSU) is really the maximum for continuous draw, and as you pointed out, there's also monitors, sound, etc. A typical monitor now is probably 50W or so (which is actually really good compared to the CRT behemoths of yesteryear).
Turn the tower on its side so the motherboard is parallel with the ground and the weight of the GPU keeps it in the PCI-e slot. It is my understanding that GPUs are able to still properly dissipate their heat in this configuration.
I have a horizontal PC chassis. Easy access to everything and has two large 140MM fans in front. It runs cool and nearly silent. I am never going back to a tower for my main PC again.
That is what I started doing. I build my desktop pc's in 4u server chassis[1] I had it vertical in a tower config until a couple months ago when I thought about the dynamics of the situation a bit and moved it to the flat orientation, as you would find it in an actual rack.
1. the cases(even the cheap ones) are better built and have better airflow than most tower cases, as a bonus, no rgb or windows.
This is the one I used on my last two builds, be aware it is a bit of a chonker.
The main reason I choose it over the others is look at all them 5 1/4 bays, the whole front of it is 5 1/4 bays. 5 1/4 bays for days. do you know how many stupid drive bays, fan controllers, sd slots, switches etc you can fit in this case... By my count all of them.
not sure, perhaps(cookies?) you are seeing a different chassis type, the one I ment to link to is "4U, 15 Bays & 8 Fans" it comes with three 3u(5.25) internal drive bays(5 drives each). honestly they suck, pull them to get access to the 5.15 bays.
At this point both the motherboard and graphics card need to be mounted on the back plate of the chassis, so that they can both use tower coolers. You can already use a PCIe extender to achieve this, but it should become the standard.
1. While I agree we're beginning to reach absurd proportions, lets really analyze the situation and think about it.
2. Are there any GPUs that actually have performed physical damage on a motherboard slot?
3. GPUs are already 2-wide by default, and some are 3-wide. 4-wide GPUs will have more support from the chassis. This seems like the simpler solution, especially since most people rarely have a 2nd add in card at all in their computers these days.
4. Perhaps the real issue is that PCIe extenders need to become a thing again, and GPUs can be placed in an anchored point elsewhere on the chassis. However, extending up to 4-wide GPUs seems more likely (because PCIe needs to get faster-and-faster. GPU-to-CPU communications is growing more and more important, so the PCIe 5 and PCIe 6 lanes are going to be harder and harder to extend out).
For now, its probably just an absurd look, but I'm not 100% convinced we have a real problem yet. For years, GPUs have drawn more power than the CPU/Motherboard combined, because GPUs perform most of the work in video games (ie: Matrix multiplication to move the list of vertices to the right location, and pixel-shaders to calculate the angle of light/shadows).
That's not a reasonable answer. Many people want to buy assembled systems, and lack the skills/inclination to do it themselves. Pre-built systems is a huge $1B+ market, and "kill the market entirely" is not an acceptable answer.
Breakage is normal in any industry, what are you suggesting as an alternative? There are companies shipping nationally and internationally with very few issues
This reply doesn't make any sense. It is a viable business model. It has total revenues across all the major players of billions of dollars per year.
And how the hell do you not ship them? You're not making any sense here. There's no alternative to not shipping them, not unless you're planning on having system builders show up individually to clients' houses and assemble PCs on the spot. That business model is way less economically efficient than simply assembling PCs centrally and accepting some breakage in shipment.
Hard foam? Regardless I'd recommend against shipping with a CPU cooler or GPU installed but some SIs seem to get away with it with self-fitting hard foam packs.
> 2. Are there any GPUs that actually have performed physical damage on a motherboard slot?
I have a MSI RTX 3070 2 fan model. It hasn't damaged my PCI-E slot (I think), but it's weight and sag causes some bending that now makes it so that my fan bearing makes a loud noise if spun up high.
My solution has been to turn my PC case so the motherboard is parallel to the ground and the GPU sticks straight up, eliminating the sag. Whisper quiet now.
If this is happening with my GPU, I shudder to imagine what it's like with other GPUs out there which are much bigger and heavier.
Yeah, I do that for a long time. Had a few accidents where I panicked until it turned out the card was just sagging and the PCIe connection was being stopped.
After it happened the 3rd time I just cleaned up a little space and put the PC lying on its side. Zero problems since then.
I've seen a lot of people ignore which screws they're using to retain their GPU
The screw should have plenty of surface area to properly secure the card. You'll still have _some_ sag, but my 3 pin 3090 doesn't sag half as badly as examples I've seen of much smaller cards
I have an EVGA 3070 and also had the sag issue. My case came with a part to support the GPU though, but I didn't realize until I solved it another way: I just doubled up those plates that you screw the GPU into so there was no way it could pivot and sag.
PCIe extenders are a thing already. Current PC case fashion trends have already influenced the inclusion of extenders and a place to vertically mount the GPU to the case off the motherboard.
GPU sag is also a bit of a silly problem for the motherboard to handle when $10 universal GPU support brackets already exist.
I have one of these for a much smaller card, mostly so that cold airflow from the floor intake fans actually has a path to get up to the RAM and VRMs. This is a workaround for a case that doesn't have front intake, which is preferable in my opinion.
It does look a little cool, but I always worry a little about the reliability of the cable itself. Does it REALLY meet the signal integrity specifications for PCI-E? Probably not. But, no unexplained crashes or glitches so far and this build is over 2 years old.
LTT has a video where they tried to see how many PCIe riser cables they could string together before it stopped working.[1] They got to several meters. Maybe you could argue that it's worse inside a PC case since there's more EMI, but it seems like your PCIe riser cable would have to be very out of spec before you'd notice anything.
The 650 Ti is PCIe 3.0. PCIe 4.0 doubles the bandwidth. PCIe 5.0 doubles the bandwidth again. The RTX 40 series GPUs still use PCIe 4.0, which have commonly available conformant riser cables. I suspect the story for PCIe 5.0 will be different.
It's not clear whether they reached a limit of drive strength or latency (I doubt EMI is the factor, since he said those are shielded) but that's a good demonstration of the resiliency of self-clocked differential serial (and aggregated serial) buses. The technology is much closer to networking standards like Ethernet than traditional parallel PCI, with features like built-in checksums, error correction (newer versions), and automatic retransmit.
I wonder if that benchmark actually loaded the PCIe bus to any significant degree after the initial benchmark startup, or just updated a couple small parameters on a single scene and thus mainly just tested the local computation on the GPU?
You'd want to somehow monitor the PCIe bus error rate - with a marginal signal and lots of errors -> retries, something that loads the bus harder (loading new textures etc) could suffer way more.
They do briefly show a different PCIe riser made out of generic ribbon cable [1, 3:27], and say that one failed after chaining only two ~200mm lengths. The quality of the riser cable certainly matters.
You need Steve for that kind of testing, LTT would be busy putting rgb on it and then (badly) water cooling it so they could sell you a backpack with no warranty.
Yeah, I have a motherboard with a bent out of shape port because of the weight of the card in it. My current desktop has a sag of easily half an inch at the end of the card and it’s not even a particularly large one by current standards. The ATX spec obviously wasn’t designed with GPUs this heavy and this power hungry.
historically cases had a bracket at the front to support full length cards. I even remember I once had a reference amd card that had an odd extension so that it would be supported by the forward full length brackets.
I have to admit I have not seen that front bracket for a long time. some server chassis have a bar across the top to support large cards. this would bet great except gfx card manufacturers like to exceed the pci spec for height. that bar had to be removed on my last two builds. now days I just put my case horizontal and pray.
I came here to mention the front support bracket. You'll find it on the larger AMD workstation cards more often than others, I first remember it on a FirePro reference card, and some searching turned up examples of it for the AMD FirePro V7900, and a few other models.
I've also had the vertical clearance issue since I try to rack mount all my gear now I've got a soundproof rack, its very annoying to need more than 4U of space just to fit a video card in a stable vertical orientation.
I’m not sure how the author, who programs GPUs, doesn’t comprehend the cooling is part of market segmentation. A 3090 Turbo is a two slot solution, but NVIDIA forced vendors to discontinue it to prevent creep into datacenters.
And I’m sure the licensing bros will come out and shout about licensing or something irrelevant. My dudes, I operate 3090s in the data center, it saves boatloads of money for both upfront, licensing, power and therefore TCO, and fuck NVIDIA.
My proposal isn’t too different. Move the one ultra-fast PCIe slot to the top of the motherboard. It would be mounted so the GPU plugs in parallel to the motherboard, on the same plane, above the motherboard on a tower case. The few other PCIe slots that exist can stay at the bottom.
Only downside is the case has to be taller. Not sure if that would be considered a problem or not.
This doesn’t really help dual GPU setups, but those have never been common. I don’t have a good solution there. I guess you’re back to some variation of the riser idea.
On a further note, why does it even have to be inside the case? Make a slit in the case on the top, so that the PCIe slot is sticking out. Stick a GPU in that slot, supported by the case. The GPUs these days look much cooler anyways.
Not necessarily. A flexible interconnect would allow the GPU to be planar with the MB; just bend it 180 degrees. Now your GPU and CPU can have coolers with good airflow instead of the farcical CPU (125mm+ tall heatsinks...) and GPU cooling designs (three fans blowing into the PCB and exhausting through little holes...) prevailing today.
My idea is to separate the whole CPU complex (the CPU and heatsink, RAM slots, VR, etc.) from everything else and use short, flexible interconnects to attach GPUs, NVMe and backplane slots to the CPU PCB, arranged in whatever scheme you wish.
I was kind of hoping doing it that way would let you put big CPU style coolers on the GPU parts with a lot more height than a 1x or 2x expansion slot.
If you “folded” the GPU over the CPU to save height I would think that would be worse than today for heat.
Maybe I’ve got this backwards. Give up on PCIe, or put it above the rest of the motherboard. The one GPU slot, planar to the motherboard, stays below. Basically my previous idea flipped vertically.
The other PCIe slots don’t need to run as fast and may be able to take the extra signal distance needed. The GPU could secure to the backplane (like my original idea) but would have tons of room for cooling like the CPU.
> If you “folded” the GPU over the CPU to save height I would think that would be worse than today for heat.
Why? Cooling would be far better: the CPU and GPU heatsinks would both face outward from the center and receive clean front-to-rear airflow. Thus, looking down from above:
The power supply and power connectors are on the bottom. Another PCB lays flat on top to host open PCI-E slots, NVMe, whatever, connected 90 deg. to the CPU PCB with one PCI-E slot interconnect. All interconnects are short. Air flow is simple and linear. The CPU/GPU Heatsinks are large and passive: you only need intake fans.
I've been refining this. I'm actually learning FreeCAD to knock out a realistic 3D model.
One obvious change: run the CPU/GPU interconnect across the bottom: existing GPU designs could be used unmodified (or enhanced with only a new heatsink) and the 16x PCI-E lanes for the GPU would be easier to route off the CPU PCB.
There are virtually no significant differences between the motherboard layout IBM promulgated with original IBM PC (model 5150) in 1981 and what we have today. That machine had a 63W power supply and no heatsinks or fans outside the power supply. The solution to all existing problems with full featured, high power desktop machines is replacing the obsolescent motherboard design with something that accommodates what people have been building for at least 20 years now (approximately since the Prescott era and the rise of high power GPUs.)
Not a bad idea however you have hard limits on how physically long the PCIe lanes can be. We had problems making sure we hit signal budget for a PCIe gen 4 slot on an ATX motherboard. The problem (PCIe lane length) gets worse as the speeds increase.
With mini cube PCs growing in popularity the future will probably be this, a mini PC with every part a USB type modularity for any ram or GPU or hd into that stock cube PC.
The main reason is that CPUs get old much slower than GPUs, but “current” socket generations change quickly. Another reason is a combinatorial explosion of mobo x graphics card features, which is already hard.
Gpus get old slowly too, but youtube influencers convince gamers their current card isn’t enough often ans effectively (I’m not excluded from this group)
The GPU is the most expensive component in gaming PCs these days, thus it makes the least sense for it to be the hard-wired component as there is the most price diversity in options. I have definitely upgraded GPUs on several computers over the past decade and I'm very thankful I didn't have to rip out the entire motherboard (and detach/reattach everything) to do so.
It's only the cheap components without a wide variety of options that make sense to build in, like WiFi, USB, and Ethernet.
This is already happening (as was noted in other comments on this article).
One of the most prominent examples is the entire Apple Silicon lineup, which has the GPU integrated as part of the SoC, and is powerful enough to drive relatively recent games. (No idea just what its limits are, but I'm quite happy with the games my M1 Max can play.)
Note this whole discussion is in the context of the 4090. If you're an enthusiast, soldering the GPU to the mobo forces you to spend $200-$700 more every time you upgrade your GPU because you also have to buy a new mobo and possibly a new CPU if the socket changed.
The GPU is also one of the easiest components to swap today. That's not something I want to give up unless I see performance improvements. Cooling doesn't count because I already have an external radiator attached to my open-frame "case".
I went through 3 GPUs before changing motherboards and I'm still bottlenecked on my 3090, not my 5800X3D. After I get a 4090, I expect to upgrade it at least once before getting a new mobo.
Having had a few GPUs go bad on me over the years, I would hate to have to disassemble the entire thing to extricate the mobo/gpu combo for RMA'ing, rather than just removing the GPU and sending it off.
Then you lose modularity, which is a huge benefit to PC gaming? Now if you want to upgrade to the newest graphics card, you also need to get a new motherboard. Which also could mean you need a new CPU, which also could mean you need new RAM.
Right now you can just switch the graphics card out for another and keep everything else the same.
> 2. Are there any GPUs that actually have performed physical damage on a motherboard slot?
Yes. I've seen both a heavy GPU and an HSM card damage a slot. One happened when a machine was shipped by a commercial shipper. The other happened when the machine was moved between residences. It doesn't occur to people that the mass of a card swinging around is a problem when the case is moved.
The HSM one was remarkable in that it was a full length card with a proper case support on both ends.
Also, this isn't just about damaging the PCI-E slot. Heavy cards bend PCBs (both the MB and the dangling card) and bending PCBs is a bad idea: surface mount devices can crack, especially MLCCs, and solder joints aren't flexible either. No telling how many unanalyzed failures happen due to this.
If you have a big GPU don't let it dangle. Support it somehow.
Another area where the conventional layout is struggling is NVME. They keep getting hotter and so the heatsinks keep getting larger. Motherboard designers are squeezing in NVME slots wherever they can, often where there is little to no airflow...
> One happened when a machine was shipped by a commercial shipper. The other happened when the machine was moved between residences. It doesn't occur to people that the mass of a card swinging around is a problem when the case is moved.
Huh. Good point. I'll be moving soon, and have kept the box my case came in as well as the foam inserts for either end of the case. I might just remove the GPU and put it back in its own box for the move, as well. Thanks for bringing that up.
> Are there any GPUs that actually have performed physical damage on a motherboard slot?
It's quite common to suffer damage from movement, especially in shipping, to the point where integrated PC manufacturers have to go to great lengths to protect the GPU in transit.
2. We crossed that point generations ago. High end GPU owners are advised to remove GPUs from their system before transporting it and PC communities get posts often enough of people who had consequences from not doing so. Over a longer term even a lighter card can deform motherboards - I had a 290x which did so to an old Z87 motherboard over 5 years with the result the motherboard was no longer flat enough to mount the backplate of a replacement cpu cooler to.
4. Don't forget high end GPUs are also getting longer, not just thicker. So increasing sizes both give and take away weight advantages
Gpu sag is a big issue in gaming computers. I had a small (by comparison and of more contemporary graphics cards) rx480 and bought a cheap 10 dollar graphics card brace to reduce its strain on the pci slot in 2021 to help reduce its chances of failure during the shortage. I use the brace to hold up my new ampere card now (which is maybe twice the length of the rx480).
Workstation cards are traditionally supported on 3 sides so don’t suffer this sag. In such systems there’s usually an extension bracket screwed into the card to increase the length, allowing it to reach a support slot to hold it steady.
It's not a motherboard problem. How would you integrate support for user provisioned cooling options (to match cooler to the card wattage) and still keep any sort of flexible expansion slot area? GPUs can't be turned into a single chip, there's too much going on, so you're never going to have a CPU cooler situation. So, fine, what if you made them daughterboards mounted like M2 SSD's; that may work, except ATX now has to become ITX to give room for an extra board's worth of space.
It's a PC case orthodoxy issue, really. People want plugs at the back of the box, which dictates how the GPU must sit in the case, and disagreement on GPU sizing means no brackets. Solve these two issues and life gets a lot better.
Or, solve it like SFF case guys solved this problem, by using a PCIE extender cable to allow the GPU to be mounted wherever you like.
Why can't I find a recent GPU that doesn't take up half my case? Do the GPU manufacturers not care about people who want more than onboard graphics but don't need the most powerful gaming GPU?
You can? Both the RTX 3050 and RX 6400 come in single fan form factors not much longer than a PCIE slot. Anything weaker than that is getting beaten by AGUs.
For reference, there used to be low profile cards available barely longer than an AGP slot. For example, here's a low profile Nvidia Gefore4 MX440 (I used to have one of these):
> Anything weaker than that is getting beaten by AGUs.
Is AGU the GPU on the CPU thing? Is it not possible (or maybe not profitable) to put something a little more performant than an AGU but doesn't need such a massive fan on a full size multiple slot card?
APU*, for accelerated processing unit. Essentially a CPU with integrated graphics, but the graphics cores can also be used for traditional CPU computing if it's not being used for graphics.
They can't put anything bigger because they run into heat, power, and memory limits.
Barely longer than the PCI-E slot and only needs a single slot.
The Radeon RX 6400 is basically AMD's high end integrated GPU on a card being very similar to the Radeon 680M in Ryzen 6000 chips. But I think the dedicated memory offers an improvement over the APU solution.
Simple: GPU manufacturers would much rather sell you a flagship at $1000+ than multiple basic cards at $150. Duopoly in the market and no competition, why would they reduce their margins?
> flagship at $1000+ than multiple basic cards at $150
Those aren't even the same market segment.
The type of person buying a flagship card would not even be considering a basic card and vice-versa.
Also, a part of me used to get frustrated at the existence of the "basic" cards, like the GT 1030. I've seen more than one person wanting to build a gaming PC and see a budget card like that and think they're getting a current-gen card without knowing that the budget cards are horrendously underpowered. For example, the GT 1030 is about the speed of a GTX 470, a mid-grade card from 7 years before it.
Maybe some people are fine getting that weak of a card, but if that's all you need, I'd question if you even need a discrete GPU.
> Or maybe they could devise some sort of stacking system, with each GPU board separate and stacked.
Yeah, I remember those stackable/modular computer concepts that industrial-design students loved to put in their portfolios from the late-1980s to the mid-1990s; I get the concept: component/modular PCs are kinda like the modular hi-fi systems of the 1980s, except those design students consistently failed to consider how coordinated aesthetics are the first thing to go out the window when the biz-dev people need the hardware folks to save costs or integrate with a third-party, etc.
...it feels like a rare miracle that at least we have 19-inch racks to fall-back on, but those are hardly beautiful (except the datacenter cable-porn, of course):
Multi-gpu setups for computation could have two SKUs one "motherboard" SKU and one connectorized SKU with the connector NOT being PCIe (after a transition).
They already do multi-form-factor, PCIe and OAM for AMD, PCIe and SXM for Nvidia.
Just drop PCIe, have a large motherboard SKU with a CPU slot and some OAM/SXM connectors in quantities up to what a wall socket can supply in terms of wattage (so, like 1 or 2 lol).
Vestigial PCIe cards can hang off the CPU card, if they're even needed.
High speed networking and storage is already moving to the DPU so these big GPUs, unhindered by the PCIe form-factor, could just integrate some DPU cores like they do RT cores, and integrate high speed networking and storage controllers into the GPU.
Home user? You get 1 or 2 DPU cores for NVMe and 10-gig Ethernet. Datacenter? 64 DPU cores for 100-gig and storage acceleration. Easy-peasy.
The motherboard and then 1 or 2 expansion sockets, for a total of 3.
3 x 450W for GPU (Nvidia says 4090 draws 450W-- I think they're lying and will wait for reviews to see the truth) and 500W for rest of system. Though that might be a bit low, the 11900K has been measured at what, 300W under load? And you'd need a budget for USB-C power, multiple NVMe, fans, and whatnot. Maybe the spec would accommodate high-core-count enterprise processors so 600W+ would be wiser.
Even in euroland 2KW, which is what a maxed out system would be at the wall socket, is a bit much. They don't even allow vacuum cleaners over 900W to be sold.
2kW isn't actually that much — you can buy ATX power supplies up to 3.5kW for usage in 230V countries. Often used for mining, but also useful if you want to accelerate your blender renders by just throwing multiple GPUs at the task.
Multi-GPU isn’t just about single-application performance.
Multi-GPU is necessary if you need more display output ports: with limited exceptions every GPU I’ve seen is 3xDP+1xHDMI or worse. While a single DP can drive multiple monitors it limits your max-res/refresh-rate/color-depth - so in-practice if you want to drive multiple 5K/6K/8K monitors or game at 120Hz+ you’ll need two or more cards, with-or-without SLI/etc.
Weirdly, this is the design of the Raspberry Pi, wherein the GPU bootstraps the system and brings the CPU up. Though it is a somewhat academic distinction, since both live inside the same chip.
Compare the size of the new air-cooled 40XX cards and the iChill 4090, which is tiny by comparison. The simple answer is just to use liquid cooling if you have a card using 400w. Then all the absurdity just goes away.
The water cooling on my 3090 is heavier than the air cooling for the same card, as you need a coldplate on both sides of the thing to get rid of the heat.
Could we just follow the atx spec? There is a maximum length for expansion cards, and at that end there are optional supports. These are in servers already. Just start using all that case volume to support the GPU.
Cheers, bravo, exactly what I wanted to say. However just to be pedantic, I just read the atx spec to find this exact information and it is not there. I suspect the expansion card dimensions are in the pci spec.
My case has these and I would love to know the proper keywords to find extension brackets I could add to my shorter cards. Searching for "extender" returns electrical extenders for the PCIe port itself.
Problem is that the card retainers in the spec would all, I suppose, need to align with the chassis. Card widths are highly variable, so all manufacturers would need to change their card designs to allow for at least providing aftermarket retainers on them.
> A 4080 series card will demand a whopping 450 W,
No, that was just a rumour that was floating around. The 4080 16GB model is 340W TGP, the 12 GB is 285W TGP out of the box. The 3080 (10 GB) was 320W TGP, as a comparison point.
Could be. But the pre-launch rumours also claimed the 4080 was a 450 W product, and the 4090 is so niche (basically it’s the old Titan line) that it’s not worth handwringing about because it’s not targeted at normal consumers (for “I’d buy a > $1000 GPU for gaming” values of “normal”).
Nvidia is attempting to increase sales on the high-end by positioning the 4080 and 4070 (aka 4080 12 GB) at a much higher price point than the market demands while giving good value to the 4090 (60% more performance than the 3090 Ti which launched at $2000).
How about we move to external only GPUs with huge connectors? If GPUs are half the size, power consumption and price of a PC now, they might as well be a separate device. As a bonus the rest of the motherboard & PCs actually get much smaller. A PC without any spinning disks could conceivably just be the size of a NUC by default, something you can travel with when you don't need the beefy GPU.
there are PCIe riser cables that you can imagine as an extension cord for a PCIe slot. they're popular in some thin server chassis or in miniITX builds where the form factor dictates the GPU must be in a certain orientation to fit.
there's also the method of external GPUs via Thunderbolt (which carries PCIe) that is more practical for putting a GPU in an entirely separate enclosure
Given that every Mac, AMD-based XBox and Playstation, and a chunk of the Ryzen market are shipping decent unified CPU/GPU combos, I would say that day is mostly here.
I’m inclined to say that’s more due to integrated gpus not being a good replacement if you need CUDA or play games; If a better combined alternative existed (and not at an outrageous price) I don’t see the majority skipping it die to lack of being modular.
The most expensive part of GPU operation (power-wise) isn't computation, it's actually moving information around
I doubt that. Compare the GPU temperature (a good proxy for power consumption) when playing a game or doing GPGPU stuff vs playing a video (without GPU acceleration, so it's just acting as a framebuffer). The former involves far more computation, and gets the GPU much hotter.
No. There was an internal nvidia presentation from a few years ago that stated that moving data was the hard part. (I can't find the presentation any more, but if anyone can find it, please post it below.)
Previously graphics cards were essentially designed with a single area of the card handling computation, and another area holding memory. Data would need to be moved from the memory, to the computation area, then back again if there were changes that needed to be stored.
As the computation and the memory demands became larger, those areas had to handle more, but so did the bus between those two areas. What was a negligible overhead for the bus became more pronounced as it had to handle more data.
Eventually the energy overhead of transporting that much data across that distance started constraining what was possible with graphics cards.
That is why graphics card architectures have shifted over time to place memory cache units next to computation units. The less distance the data needs to travel, the smaller the power requirements. It's also led to the investment and adoption of stacked memory dies (why grab data from 5cm away in the x-y plane when we can stack memory and grab data 5mm away in the z-direction).
Moving around data is indeed a major issue for any throughout oriented device. But for a gaming GPU, PCIe BW has never been an issue in any of the benchmarks that I’ve seen. (Those benchmarks artificially reduce the number of PCIe lanes.)
In fact, the 4000 series still has PCIe 4.
Moving data around for a GPU is about feeding the shader cores by the memory system. PCIe is way too slow to make that happen. That’s why a GPU has gigabytes of local RAM.
The problem with external gpus is that the pcie5x16 slot requires signal integrity chips every 5 inches or so [1]. Even for pcie4 it's bad, many people had blue screens or boot issues when using pcie3 only rated riser cables with a pcie4 video card, even though electrically they have the same connections. So, having a huge cable with x16 lanes of pcie5 on the back of the computer doesn't seem that feasible. Maybe fiber optics could be a thing.
Fiber optic would have higher latency vs. copper. Not sure how much of a difference that would make on a bus but I would assume the timing is pretty tight.
I suppose it's possible, but I have yet to see an actual GPU that couldn't be fixed by properly supporting the back bracket, as shown in this video. https://www.youtube.com/watch?v=liChy76nQJ4&t=591 (Starts around 10:00 in)
I predict the GPU will need to be externalized in the future with its own cooling and power system. You'll plug into the motherboard via some cable interconnect.
They're simply getting too big, power hungry and hot to keep in colocated in the case.
Power grids are not buckling everywhere. Also this card uses less than a third of a single space heater. Finally, this card has not been released yet, so it isn't causing anything.
According to the Steam Hardware Survey (which you can view for yourself at <https://store.steampowered.com/hwsurvey/Steam-Hardware-Softw...>), the RTX 3090 makes up less than 0.48% of all video cards in use by gamers. The 450W power figure is for the new RTX 4090, which will no doubt take a while to reach even that level.
It might be nice to have more information than that, but it honestly doesn’t seem like a huge problem. Your fridge probably costs you more over a year than even a 4090. See also air conditioning.
I'd rather have a dedicated socket on the motherboard where I can just slap in a GPU chip and use whatever RAM I have soldered onto the mobo around the GPU socket.
I wonder whether pcie cabling (like oculink) will scale/work with newer pcie specs.
I have long thought the bitcoin miners were onto something, with pcie risers galore. In my head I know pcb is cheap and connectors - cables arent but it always seemed so tempting anyways; very small boards, cpu & memory (or onchip memory) & vrm, and then just pipes to peripherals & network (and with specs like CXL 3.0, we kind of could be getting both at once).
This specific problem can be solved by rotating motherboard 90 degrees (there are a few Silverstone cases that are laid out like this, they also tend to have excellent air cooling performance).
Yeah, I've had a Silverstone Raven for 10+ years, so I'm not sure why those types of cases haven't become the standard. They have their flaws, like it being awkward to route your cables into the top area, and needing good airspace underneath the case. But it's balanced out by being able to easily plug stuff into the top when you don't care about pretty cable routing, a focus on powerful, adjustable airflow, and less need for space on the sides. I'm guessing vertical space is the main problem for people.
What if we just have blocks of quartz and use laser arrays to build photonic switching junctions, no more cooling problems because it's just photons ¯\(°_o)/¯
Seriously though, I imagine it's only a matter of time before these engineering decisions are themselves handed off to machines.
The technology is getting smaller! It's just that at the same time, people still demand orders of magnitude higher performance. There are physical limits to how small a cooler can be, and that takes up like 95% of the space.
361 comments
[ 3.5 ms ] story [ 274 ms ] threadBut yea, I have to say peak power consumption has to be regulated so companies compete in efficiency, not raw power.
Yes, PEAK instantaneous wattage could easily be well over 1kW for a crazy rig like that, but unless you are specifically torture testing the rig over a long duration, it simply isn’t accurate that you are drawing anywhere near the max continuously under gaming conditions.
> Two 3090s plus a 64 core Threadripper pro will get you beyond 1000 watts alone
3090s are 320/350W each and 64 core Threadripper Pro has a TDP of 280W, just as an FYI.
Show me an example of a “non-crazy” build that’s over 1,200W, let alone over 1,500W. I’m genuinely curious.
Plus other things on the circuit. Fancy ultra bright HDR monitors use a decent chunk of power, and anything else.
Per NEC in the US, the wire used in a circuit has to be rated to handle a current draw of up to 80% continuously at the specified temperature max for the type of installation.
80% of 15 is 12, the average home in the US has split phase incoming power with a nominal voltage of 120V (1440W) with many as high as 125V (1500W). Note that is current draw over long periods of time, which a gaming rig under heavy gaming won’t usually do. Maybe a mining rig back in the day would, but otherwise it’s just not true you’re continuously drawing anywhere near 1,000+ watts on anything but the most crazy setups.
> I said near, we still have some breathing room.
No, you made a wild ass conjecture not based in fact and when gently questioned on it, you shifted goal posts and made up more conjecture.
I have an i9-9900K and an RTX 3080. My computer is also plugged into a Kill-a-watt to measure its power usage.
I ran Prime95's CPU torture test using the max power option while also crypto-mining on the GPU. I peaked at 650 watts for a couple seconds until the CPU began thermal throttling.
If I can't hit 700 watts while trying to use as much energy as possible, I can't imagine what monster system would even touch 1000 watts, assuming we're still talking consumer-grade.
So yeah, there's plenty of head space but I have also hit circuit breaker issues recently on a proper (non consumer) workstation so the limit is fresh in my mind.
For continuous draw you’re limited to 80% of the line’s rating (in the US). People who have looked into EV charging have certainly run into this fact.
So the computer (and sound/monitor/etc) have to stay under ~1400 watts if it’s on a dedicated circuit.
1/3rd of that for just the GPU I’d pretty crazy. I honestly wouldn’t have thought GPUs could get this high.
Also, I remember my computers being hot in a small room long ago. A lot of the space heaters on Amazon are rated either 750 or 1500 W.
I don’t think I wanna be near a computer that hot.
Great article!
Less sarcastically, I would quite like to go back to putting my monitor on top of a desktop - sure beats a stack of books.
People could also move to a rack mountable chassis.
1. the cases(even the cheap ones) are better built and have better airflow than most tower cases, as a bonus, no rgb or windows.
The main reason I choose it over the others is look at all them 5 1/4 bays, the whole front of it is 5 1/4 bays. 5 1/4 bays for days. do you know how many stupid drive bays, fan controllers, sd slots, switches etc you can fit in this case... By my count all of them.
https://www.newegg.com/rosewill-rsv-l4500u-black/p/N82E16811...
https://www.newegg.com/Controller-Panels/SubCategory/ID-11
2. Are there any GPUs that actually have performed physical damage on a motherboard slot?
3. GPUs are already 2-wide by default, and some are 3-wide. 4-wide GPUs will have more support from the chassis. This seems like the simpler solution, especially since most people rarely have a 2nd add in card at all in their computers these days.
4. Perhaps the real issue is that PCIe extenders need to become a thing again, and GPUs can be placed in an anchored point elsewhere on the chassis. However, extending up to 4-wide GPUs seems more likely (because PCIe needs to get faster-and-faster. GPU-to-CPU communications is growing more and more important, so the PCIe 5 and PCIe 6 lanes are going to be harder and harder to extend out).
For now, its probably just an absurd look, but I'm not 100% convinced we have a real problem yet. For years, GPUs have drawn more power than the CPU/Motherboard combined, because GPUs perform most of the work in video games (ie: Matrix multiplication to move the list of vertices to the right location, and pixel-shaders to calculate the angle of light/shadows).
I said don't ship them. I did not say don't build prebuilt systems.
And how the hell do you not ship them? You're not making any sense here. There's no alternative to not shipping them, not unless you're planning on having system builders show up individually to clients' houses and assemble PCs on the spot. That business model is way less economically efficient than simply assembling PCs centrally and accepting some breakage in shipment.
I have a MSI RTX 3070 2 fan model. It hasn't damaged my PCI-E slot (I think), but it's weight and sag causes some bending that now makes it so that my fan bearing makes a loud noise if spun up high.
My solution has been to turn my PC case so the motherboard is parallel to the ground and the GPU sticks straight up, eliminating the sag. Whisper quiet now.
If this is happening with my GPU, I shudder to imagine what it's like with other GPUs out there which are much bigger and heavier.
After it happened the 3rd time I just cleaned up a little space and put the PC lying on its side. Zero problems since then.
The screw should have plenty of surface area to properly secure the card. You'll still have _some_ sag, but my 3 pin 3090 doesn't sag half as badly as examples I've seen of much smaller cards
PCIe extenders are a thing already. Current PC case fashion trends have already influenced the inclusion of extenders and a place to vertically mount the GPU to the case off the motherboard.
GPU sag is also a bit of a silly problem for the motherboard to handle when $10 universal GPU support brackets already exist.
It does look a little cool, but I always worry a little about the reliability of the cable itself. Does it REALLY meet the signal integrity specifications for PCI-E? Probably not. But, no unexplained crashes or glitches so far and this build is over 2 years old.
[1] https://www.youtube.com/watch?v=q5xvwPa3r7M
The 650 Ti is PCIe 3.0. PCIe 4.0 doubles the bandwidth. PCIe 5.0 doubles the bandwidth again. The RTX 40 series GPUs still use PCIe 4.0, which have commonly available conformant riser cables. I suspect the story for PCIe 5.0 will be different.
You'd want to somehow monitor the PCIe bus error rate - with a marginal signal and lots of errors -> retries, something that loads the bus harder (loading new textures etc) could suffer way more.
They do briefly show a different PCIe riser made out of generic ribbon cable [1, 3:27], and say that one failed after chaining only two ~200mm lengths. The quality of the riser cable certainly matters.
[1] https://youtu.be/q5xvwPa3r7M?t=207
I have to admit I have not seen that front bracket for a long time. some server chassis have a bar across the top to support large cards. this would bet great except gfx card manufacturers like to exceed the pci spec for height. that bar had to be removed on my last two builds. now days I just put my case horizontal and pray.
I've also had the vertical clearance issue since I try to rack mount all my gear now I've got a soundproof rack, its very annoying to need more than 4U of space just to fit a video card in a stable vertical orientation.
And I’m sure the licensing bros will come out and shout about licensing or something irrelevant. My dudes, I operate 3090s in the data center, it saves boatloads of money for both upfront, licensing, power and therefore TCO, and fuck NVIDIA.
Only downside is the case has to be taller. Not sure if that would be considered a problem or not.
This doesn’t really help dual GPU setups, but those have never been common. I don’t have a good solution there. I guess you’re back to some variation of the riser idea.
Not necessarily. A flexible interconnect would allow the GPU to be planar with the MB; just bend it 180 degrees. Now your GPU and CPU can have coolers with good airflow instead of the farcical CPU (125mm+ tall heatsinks...) and GPU cooling designs (three fans blowing into the PCB and exhausting through little holes...) prevailing today.
My idea is to separate the whole CPU complex (the CPU and heatsink, RAM slots, VR, etc.) from everything else and use short, flexible interconnects to attach GPUs, NVMe and backplane slots to the CPU PCB, arranged in whatever scheme you wish.
If you “folded” the GPU over the CPU to save height I would think that would be worse than today for heat.
Maybe I’ve got this backwards. Give up on PCIe, or put it above the rest of the motherboard. The one GPU slot, planar to the motherboard, stays below. Basically my previous idea flipped vertically.
The other PCIe slots don’t need to run as fast and may be able to take the extra signal distance needed. The GPU could secure to the backplane (like my original idea) but would have tons of room for cooling like the CPU.
Why? Cooling would be far better: the CPU and GPU heatsinks would both face outward from the center and receive clean front-to-rear airflow. Thus, looking down from above:
The power supply and power connectors are on the bottom. Another PCB lays flat on top to host open PCI-E slots, NVMe, whatever, connected 90 deg. to the CPU PCB with one PCI-E slot interconnect. All interconnects are short. Air flow is simple and linear. The CPU/GPU Heatsinks are large and passive: you only need intake fans.I've been refining this. I'm actually learning FreeCAD to knock out a realistic 3D model.
One obvious change: run the CPU/GPU interconnect across the bottom: existing GPU designs could be used unmodified (or enhanced with only a new heatsink) and the 16x PCI-E lanes for the GPU would be easier to route off the CPU PCB.
There are virtually no significant differences between the motherboard layout IBM promulgated with original IBM PC (model 5150) in 1981 and what we have today. That machine had a 63W power supply and no heatsinks or fans outside the power supply. The solution to all existing problems with full featured, high power desktop machines is replacing the obsolescent motherboard design with something that accommodates what people have been building for at least 20 years now (approximately since the Prescott era and the rise of high power GPUs.)
Thinking about it though isn’t that where a lot of motherboards put a lot of their power regulation circuitry? I’m sure something would have to move.
But excellent point. That could easily bet a much bigger sticking point that I didn’t really consider too much.
They are with the people who spend the most money on GPUs
Seems with the last GPU generation most enthusiast websites recommend to go for a bigger card instead of going SLI.
Just star designing motherboards with integrated GPUs of sufficient size.
Case closed.
It's only the cheap components without a wide variety of options that make sense to build in, like WiFi, USB, and Ethernet.
One of the most prominent examples is the entire Apple Silicon lineup, which has the GPU integrated as part of the SoC, and is powerful enough to drive relatively recent games. (No idea just what its limits are, but I'm quite happy with the games my M1 Max can play.)
The GPU is also one of the easiest components to swap today. That's not something I want to give up unless I see performance improvements. Cooling doesn't count because I already have an external radiator attached to my open-frame "case".
I went through 3 GPUs before changing motherboards and I'm still bottlenecked on my 3090, not my 5800X3D. After I get a 4090, I expect to upgrade it at least once before getting a new mobo.
Right now you can just switch the graphics card out for another and keep everything else the same.
Yes. I've seen both a heavy GPU and an HSM card damage a slot. One happened when a machine was shipped by a commercial shipper. The other happened when the machine was moved between residences. It doesn't occur to people that the mass of a card swinging around is a problem when the case is moved.
The HSM one was remarkable in that it was a full length card with a proper case support on both ends.
Also, this isn't just about damaging the PCI-E slot. Heavy cards bend PCBs (both the MB and the dangling card) and bending PCBs is a bad idea: surface mount devices can crack, especially MLCCs, and solder joints aren't flexible either. No telling how many unanalyzed failures happen due to this.
If you have a big GPU don't let it dangle. Support it somehow.
Another area where the conventional layout is struggling is NVME. They keep getting hotter and so the heatsinks keep getting larger. Motherboard designers are squeezing in NVME slots wherever they can, often where there is little to no airflow...
Or remove it. Takes five minutes to unplug and unscrew a card. Had to do that with my 6900xt recently while moving.
Huh. Good point. I'll be moving soon, and have kept the box my case came in as well as the foam inserts for either end of the case. I might just remove the GPU and put it back in its own box for the move, as well. Thanks for bringing that up.
It's quite common to suffer damage from movement, especially in shipping, to the point where integrated PC manufacturers have to go to great lengths to protect the GPU in transit.
4. Don't forget high end GPUs are also getting longer, not just thicker. So increasing sizes both give and take away weight advantages
It's a PC case orthodoxy issue, really. People want plugs at the back of the box, which dictates how the GPU must sit in the case, and disagreement on GPU sizing means no brackets. Solve these two issues and life gets a lot better.
Or, solve it like SFF case guys solved this problem, by using a PCIE extender cable to allow the GPU to be mounted wherever you like.
https://c1.neweggimages.com/NeweggImage/ProductImage/14-135-...
> Anything weaker than that is getting beaten by AGUs.
Is AGU the GPU on the CPU thing? Is it not possible (or maybe not profitable) to put something a little more performant than an AGU but doesn't need such a massive fan on a full size multiple slot card?
They can't put anything bigger because they run into heat, power, and memory limits.
https://www.powercolor.com/product?id=1640245090
Barely longer than the PCI-E slot and only needs a single slot.
The Radeon RX 6400 is basically AMD's high end integrated GPU on a card being very similar to the Radeon 680M in Ryzen 6000 chips. But I think the dedicated memory offers an improvement over the APU solution.
Those aren't even the same market segment.
The type of person buying a flagship card would not even be considering a basic card and vice-versa.
Also, a part of me used to get frustrated at the existence of the "basic" cards, like the GT 1030. I've seen more than one person wanting to build a gaming PC and see a budget card like that and think they're getting a current-gen card without knowing that the budget cards are horrendously underpowered. For example, the GT 1030 is about the speed of a GTX 470, a mid-grade card from 7 years before it.
Maybe some people are fine getting that weak of a card, but if that's all you need, I'd question if you even need a discrete GPU.
Edit: Perhaps I use the term incorrectly, I mean they take up two case slots even if only a single pcie slot.
Single slot, half-profile, small fan, shorter than ATX / mATX?
This is a mini-card that'd fit in an ITX case (170mm)...
Or maybe they could devise some sort of stacking system, with each GPU board separate and stacked.
Yeah, I remember those stackable/modular computer concepts that industrial-design students loved to put in their portfolios from the late-1980s to the mid-1990s; I get the concept: component/modular PCs are kinda like the modular hi-fi systems of the 1980s, except those design students consistently failed to consider how coordinated aesthetics are the first thing to go out the window when the biz-dev people need the hardware folks to save costs or integrate with a third-party, etc.
...it feels like a rare miracle that at least we have 19-inch racks to fall-back on, but those are hardly beautiful (except the datacenter cable-porn, of course):
https://www.reddit.com/r/RetroFuturism/comments/gioqrp/sovie...
https://en.wikipedia.org/wiki/Sphinx_%28home_automation_syst...
Multi-gpu setups for computation could have two SKUs one "motherboard" SKU and one connectorized SKU with the connector NOT being PCIe (after a transition).
They already do multi-form-factor, PCIe and OAM for AMD, PCIe and SXM for Nvidia.
Just drop PCIe, have a large motherboard SKU with a CPU slot and some OAM/SXM connectors in quantities up to what a wall socket can supply in terms of wattage (so, like 1 or 2 lol).
Vestigial PCIe cards can hang off the CPU card, if they're even needed.
High speed networking and storage is already moving to the DPU so these big GPUs, unhindered by the PCIe form-factor, could just integrate some DPU cores like they do RT cores, and integrate high speed networking and storage controllers into the GPU.
Home user? You get 1 or 2 DPU cores for NVMe and 10-gig Ethernet. Datacenter? 64 DPU cores for 100-gig and storage acceleration. Easy-peasy.
So, you expect GPUs to take 1500W of power, each? (230V @ 16A)
The motherboard and then 1 or 2 expansion sockets, for a total of 3.
3 x 450W for GPU (Nvidia says 4090 draws 450W-- I think they're lying and will wait for reviews to see the truth) and 500W for rest of system. Though that might be a bit low, the 11900K has been measured at what, 300W under load? And you'd need a budget for USB-C power, multiple NVMe, fans, and whatnot. Maybe the spec would accommodate high-core-count enterprise processors so 600W+ would be wiser.
Even in euroland 2KW, which is what a maxed out system would be at the wall socket, is a bit much. They don't even allow vacuum cleaners over 900W to be sold.
Multi-GPU is necessary if you need more display output ports: with limited exceptions every GPU I’ve seen is 3xDP+1xHDMI or worse. While a single DP can drive multiple monitors it limits your max-res/refresh-rate/color-depth - so in-practice if you want to drive multiple 5K/6K/8K monitors or game at 120Hz+ you’ll need two or more cards, with-or-without SLI/etc.
The length of the 3090 with air cooling is what makes it so heavy on the PCIE connector - localized coldplate weight is not an issue.
and once you have liquid cooling, maybe you can go to 800w...
But these aren’t just long cards. They have huge heavy chunks of metal and fans on them in the name of cooling.
Are those brackets strong enough for the job? I remember them being basically shallow plastic slots.
No, that was just a rumour that was floating around. The 4080 16GB model is 340W TGP, the 12 GB is 285W TGP out of the box. The 3080 (10 GB) was 320W TGP, as a comparison point.
there's also the method of external GPUs via Thunderbolt (which carries PCIe) that is more practical for putting a GPU in an entirely separate enclosure
The limit is around 3 meters.
https://www.youtube.com/watch?v=q5xvwPa3r7M
Moving the GPU further away from the CPU and storage is going to lead to worse latency and power requirements.
Over the next 10-15 years, we'll probably see CPU+GPU packages become mainstream instead of CPU (with basic integrated graphics) and a separate GPU.
its already the case with arm socs.
I doubt that. Compare the GPU temperature (a good proxy for power consumption) when playing a game or doing GPGPU stuff vs playing a video (without GPU acceleration, so it's just acting as a framebuffer). The former involves far more computation, and gets the GPU much hotter.
Previously graphics cards were essentially designed with a single area of the card handling computation, and another area holding memory. Data would need to be moved from the memory, to the computation area, then back again if there were changes that needed to be stored.
As the computation and the memory demands became larger, those areas had to handle more, but so did the bus between those two areas. What was a negligible overhead for the bus became more pronounced as it had to handle more data.
Eventually the energy overhead of transporting that much data across that distance started constraining what was possible with graphics cards.
That is why graphics card architectures have shifted over time to place memory cache units next to computation units. The less distance the data needs to travel, the smaller the power requirements. It's also led to the investment and adoption of stacked memory dies (why grab data from 5cm away in the x-y plane when we can stack memory and grab data 5mm away in the z-direction).
In fact, the 4000 series still has PCIe 4.
Moving data around for a GPU is about feeding the shader cores by the memory system. PCIe is way too slow to make that happen. That’s why a GPU has gigabytes of local RAM.
[1] https://www.planetanalog.com/signal-integrity-challenges-set...
They're simply getting too big, power hungry and hot to keep in colocated in the case.
It might be nice to have more information than that, but it honestly doesn’t seem like a huge problem. Your fridge probably costs you more over a year than even a 4090. See also air conditioning.
See Texas.
Basically we need to move away from slots!
I have long thought the bitcoin miners were onto something, with pcie risers galore. In my head I know pcb is cheap and connectors - cables arent but it always seemed so tempting anyways; very small boards, cpu & memory (or onchip memory) & vrm, and then just pipes to peripherals & network (and with specs like CXL 3.0, we kind of could be getting both at once).
Seriously though, I imagine it's only a matter of time before these engineering decisions are themselves handed off to machines.
I really like the Mini ITX form factor and have multiple PCs that fit on IKEA shelves.
I prefer shorter/smaller GPUs. Better airflow. Easier at build time too.
My EVGA RTX 3050 is about the right balance.