So why cards?

[edward78]

You know the Motherboards with 2 CPU sockets, why not 1 withh a GPU socket (It would have to be Universal of course)? Just sell VRam & they would go in the slots. If it is more complex, tell me.

 

[El_Capitan]

Mostly because the GPU would have to be soldered in. If they made GPU's with pins and IHS's like CPU's, it could be a thing, but no one's done it, yet. I don't see why it couldn't be done.

 

[edward78]

Mostly because the GPU would have to be soldered in. If they made GPU's with pins and IHS's like CPU's, it could be a thing, but no one's done it, yet. I don't see why it couldn't be done.

I always thought so. It would be so nice too. Every home system could be a SFF. Like a NES size case might do it. Well prob. a little bigger. Wsih I had some influence with the right people. My brother in law use to work at a Apple store, but they would just patent it so that is out. I mean, ya we have APUs & iGPUs, but what if you are good with the CPU you have, or can't fiind 1 wuth the GPU you want?

 

[Eulogy]

People already complain about needing a new motherboard to jump CPU generations. Can you imagine having to deal with the complexity of
Intel + nVidia
Intel + AMD
AMD + nVidia (uhh)
AMD + AMD

matrix of boards and compatibility?

 

[LukeTbk]

You know the Motherboards with 2 CPU sockets, why not 1 withh a GPU socket (It would have to be Universal of course)? Just sell VRam & they would go in the slots. If it is more complex, tell me.

memory bandwith is a big one would be my guess, large amd APU could be the limit, 256 bits+ bus of gddr6 bandwith without soldering both the gpu and the memory is probably really hard to do.

With sockets and pins, every one of them become tiny antenna adding noise and with socket-memory slot you add distance has well, with the size/frequency gpu memory bus operate in you would need extreme high quality and size for socket to work.

A simple 5060 gpu has like 4.5 time the memory bandwith than a dual channel ddrt5-6000 9950x, epyc type system can do 5060 bandwith but they cost a lot with massive socket and expensive board versus an entry level gpu not the high end one, something that would support 5080 level would become quickly impraticable.

 

[cpufrost]

Not to mention power delivery. Think MSI Godlike boards are expensive?
A drop in GPU may be cheap but the price of the board itself would be crazy high.
Reminds me of the days of sockets/chipsets that worked on AMD/Intel/Cyrix/WinChip/ and others.

 

[Nobu]

They used to have the graphics on one of the chipsets, nvidia's geforce mobile graphics was pretty badass...for the time.

But GPUs get hella hot, hotter than CPUs (well, used to be), and they need a lot of memory (even still). There's just not enough real estate on most boards to accommodate all that. And if you wanted upgradable vram, that takes more space and adds latency...

 

[edward78]

They used to have the graphics on one of the chipsets, nvidia's geforce mobile graphics was pretty badass...for the time.

But GPUs get hella hot, hotter than CPUs (well, used to be), and they need a lot of memory (even still). There's just not enough real estate on most boards to accommodate all that. And if you wanted upgradable vram, that takes more space and adds latency...

Oh, ok. Would a little slot like te Pentium 2 went in be better for the GPU? The VRam would have slots like the Ram, so that is already good. So Nvidia's Arm will still need a bug fan for the RTX 5090 GPU, right?

 

[Nobu]

If you put it on a card, might as well use PCIe since it's already there, standard, and very fast. And since you have a card, might as well put memory on it.

Yes, most arm solutions solder everything to the board and have a big cooler that covers everything, I would expect a similar solution for nvidia/arm. There are some recent entries with PCIe slots, but they're more like dev boards than something you would really want as a PC/game system.

 

[LukeTbk]

Oh, ok. Would a little slot like te Pentium 2 went in be better for the GPU? The VRam would have slots like the Ram, so that is already good. So Nvidia's Arm will still need a bug fan for the RTX 5090 GPU, right?

Not soldered in vs soldered in is a big difference regardless of the socket format I think.

If an Nvidia arm cpu (or mediatek or Ampere arm cpu) use a 450w-600w 5090 gpu and if you add more power-heat by not having the ram soldered and farther away from the gpu, it will need even more cooling than what we see, the ISA used by the cpu will be irrelevant.

Yes, most arm solutions solder everything to the board and have a big cooler that covers everything, I would expect a similar solution for nvidia/arm.

It is more who use and for what they use ARM (apple-nintendo-nvidia, etc...they love coherent memory) than anything inherent, they can look like that, as regular computer board for regular computer has well:

Those nvidia with many gpu rack will also have individual gpu all individually cooled and are not soldered in for yield, ability to swap one that fail later on, the level of thermal expension tolerance needed and so on, they used socket as mentionned in the first message, on a very expensive board, with really giant socket (probably ~4000 pins type for the 1000-1200 watt and amount of signal and that possible because the HBM is direct on the gpu package and not on the motherboard):

And reversly ps5-xbox will create coherent giant soc system using x86 cpus by spending the time and money to make a custom version

 

[edward78]

Isn't this idea mostly what a iGPU/APU already is only not combined with a CPU & using Sys Ram. Why would this be any harder?

 

[LukeTbk]

Isn't this idea mostly what a iGPU/APU already is only not combined with a CPU & using Sys Ram. Why would this be any harder?

It would not be if you do it under those limitation of relatively slow gpu bandwith and relatively low power, that were all the challenge of discrete level GPU in socket would come from.

i.e. we already have gpu in socket that use motherboard ram outhere, those AMD APU, we already have high power socket gpu if the vram is on package (like the picture above), socket is giant and motherboard really not cheap (and it does not use system ram).

 

[edward78]

It would not be if you do it under those limitation of relatively slow gpu bandwith and relatively low power, that were all the challenge of discrete level GPU in socket would come from.

i.e. we already have gpu in socket that use motherboard ram outhere, those AMD APU, we already have high power socket gpu if the vram is on package (like the picture above), socket is giant and motherboard really not cheap (and it does not use system ram).

IGPUs/APUs do use Sys Ram, I said my idea would not. Maybe the wording was confusing in my post above???
Isn't this idea mostly what a iGPU/APU already is only "not" combined with a CPU & using Sys Ram. Why would this be any harder?

 

[Absalom]

Isn't this idea mostly what a iGPU/APU already is only not combined with a CPU & uses Sys Ram. Why would this be any harder?

This. If you're going to do it, do it right, don't half-ass it. Put everything on the same die or same SoC package. Shortest traces, no need for a PCIe bus, etc.

Which sounds nice and all until you recognize the technical hurdles even that brings. Such as the physical size of a 5000 series gpu die dwarfing any CPU die around today. Not to mention the amount of heat all that will produce in such a small space.

The best we can hope for is having both the CPU and GPU relatively close to each other with some dedicated bus between them, much like how consoles are designed. The shorter the traces, the better. Then there's the whole VRAM vs system ram problem. Do you share VRAM as system ram like the PS5? Basically, this problem turns into something consoles have already solved.

Then there's power delivery for all that. Now the motherboard takes on the role of power delivery of the entire system. Falure point becomes a single failure point. If something dies, you basically RMA the entire system.

It would also kill a market segment overnight, because if any of that is soldered in then motherboards now play a completely different role. Less emphasis on horizontal integration and more on vertical integration (i.e. your purchase decision comes down to motherboard + CPU + GPU combination). Going vertical makes things more like Apple hardware. Choices are limited to what configurations are provided at purchase time. Also, good luck getting X and Y companies to agree on what combos "sell". AMD is in a far better position vs getting Nvidia and Intel to agree on things.

I think what's considered "legacy" is what's keeping us tethered to the past. That and the whole "If it ain't broke, don't fix it" mentality.

 

[LukeTbk]

Isn't this idea mostly what a iGPU/APU already is only "not" combined with a CPU & using Sys Ram. Why would this be any harder?

well you had one signal that apu does not have for the board to handle, between cpu and gpu and now the gpu need its own memory controller to talk to its own memory and its own memory slot on the board, way more costly if not harder (which maybe are the same word here, it is all about cost, because there is motherboard with on socket gpu as in the picture I posted, as long as the vram unlike in the op is on gpu package you can have regular gpu-memory speed)

APU/iGPU are low power and low memory bandwiith gpu, those would be possible on a very big and costly motherboard to be as the OP describe (you add a giant socket, whole new rams signals....), but that would be a strange market, way more expensive than an AMD APU, way less performance and not that more swapable than a pci-e GPU and more expensive than those on the initial purchase, what would be the point.

 

[littlegreenmenfromhome]

I'd rather have SXM AIB.

 

[TrailRunner]

Oh, ok. Would a little slot like te Pentium 2 went in be better for the GPU? The VRam would have slots like the Ram, so that is already good. So Nvidia's Arm will still need a bug fan for the RTX 5090 GPU, right?

I think that you're getting somewhere with that packaging idea similar to the Pentium 2.
As mentioned by prior posts, you still can't have discrete RAM for a high-performance video chip, it needs to be soldered or located close to the package. For example, for integrated GPUs similar to how the Apple M processors have packaged memory, and AMD Strix Halo requires soldered ram for technical reasons. GPUs just need some fast memory closer to the GPU than is possible with replaceable RAM.

The necessity for very fast memory (much faster than traditional system ram) is what the Intel engineers encountered when designing the Pentium 2. Before the Pentium 2, system cache was located on the mainboard, and ran at the speed of the front side bus. Intel realized that their Pentium 2 processor would need soldered memory close to the processor, but doing this in the traditional socketed way would limit upgradability. To speed up the memory access of the Pentium 2, Intel packaged the processor together with some very fast cache memory onto the slot-style package. So the downside of the Pentium 2 slot packaging was that you couldn't upgrade only the CPU (as the chip was soldered to the slot-style package), instead the packaging meant that when doing a CPU upgrade you replaced the system cache at the same time (as part of the slot package). The upside is that processors and cache were developing faster concurrently, so when you upgraded the packaged slot processor, you benefited from new faster cache memory as well. So the slot packaging allowed for the benefits of processor upgrades, while also allowing for the benefits of the faster soldered cache memory that the processor required.

So, it could be a good idea to package a GPU and some fast video memory as close as possible, in some sort of slot packaging, similar to the Pentium 2 idea. When the GPU advances occur, and you replace the slot package to perform a GPU upgrade, you would benefit from advances in faster memory at the same time.
Another benefit of some sort of slot packaging, (a card if you want to think of it like that) could be that if the GPU consumes more power than the mainboard could otherwise deliver, you could have some dedicated power cabling in the system deliver additional power to the GPU. You can't do that with a socketed design.
If you think about it, you could even have the flexibility to place the slot package close to the edge of the mainboard so that it could exhaust hot air directly *outside* of the case instead of dumping it into the case itself. An additional benefit of this is that instead of routing the video signal through the slot package and through the mainboard, the system could be designed so you could plug your monitor's cables directly into a video port at the edge of the slot that faces the outside of the PC's case. So that way when DisplayPort/HDMI specifications get updated, you can just upgrade your packaged GPU instead of replacing the whole mainboard.

You could make a vendor-agnostic slot standard so GPUs from Nvidia, AMD, Intel, or any other manufacturer could be used in any mainboard. Since the packaging is for faster video, you could call it Accelerated Graphics Packaging. Or Accelerated Graphics Port. Or AGP.

Just a thought.

 

[zandor]

The way higher end cards are going I kinda wonder if maybe they ought to be more like a PSU. Something you can bolt onto your case to deal with the weight.