TR Forums

I apologize if this has been asked a thousand times, but what prevents a a manufacturer from just sticking a discrete CPU and discrete GPU on the same PCB? Is it technical or cost or flexibility or space? Maybe a combination of?

WhatMeWorry wrote:

I apologize if this has been asked a thousand times, but what prevents a a manufacturer from just sticking a discrete CPU and discrete GPU on the same PCB? Is it technical or cost or flexibility or space? Maybe a combination of?

It's called a laptop.

It's what some laptops do.

There's no technical reason why it couldn't be done. But what would be the point? Sure you'd save a little space and cost, but you'd be giving up the flexibility of a discrete GPU, and causing additional headaches from a cooling standpoint.

What are you thinking the target market would be? SFF systems?

See: laptop motherboard as a prime example, they already exist. I assume you're referring to desktop motherboards. Space, cost, and heat are the major drawbacks. In the age of IGPs, you're going to need a pretty decent dGPU to make it worthwhile. Besides space/cost/heat, the inability to upgrade is a big problem.

A discrete GPU on a motherboard could easily be called an "onboard video chipset." Sure, the "onboard video chipsets" of the not too distant past (and still lingering) are terribly anemic compared to what one would call a "discrete GPU." But they are very similar after you put them on the motherboard.

Yeah you get the odd desktop one (eg http://www.zotac.com/aa/products/graphi ... eries.html ), however the thermal solution for any decently powerful GPU would probably take up too much of the motherboard. The heatsink on the mobo I linked is just for a 6200, and it already dominates the board .

I don't have numbers for recent products, but high-performance GPUs have commonly lived on 8-12+ layer PCBs while conventional desktop motherboards have traditionally been in the range of 4-6 layers. Some of this is a function of the extremely wide, high-speed memory paths that GPUs require. Laptops do it anyway because space is more valuable than absolute manufacturing cost of the components. In the desktop arena, some of the problems associated with layout can be reduced when the CPU has an on-board memory controller and the GPU lives inside the CPU package and shares resources a la Intel Sandy/Ivy and AMD Fusion.

Another problem is that high-performance GPUs tend to produce heat in the same range as a high-performance CPU and tend to have similar cooling requirements, while BGA-packaged chips (including GPUs and chipsets) have a tendency to pop away from a PCB if it flexes too much. The additional PCB layers on a GPU card help mitigate some of this, and the form factor allows the typical cooler to be made longer and thinner. On a motherboard, the GPU cooler might require some of the same backplane reinforcement typically used around the CPU socket, which causes layout headaches and increases the cost.

Also, you can't upgrade an onboard GPU, except by installing an add-on card, which has its own economic issues.

ludi wrote:

I don't have numbers for recent products, but high-performance GPUs have commonly lived on 8-12+ layer PCBs while conventional desktop motherboards have traditionally been in the range of 4-6 layers.

Another problem is that high-performance GPUs tend to produce heat in the same range as a high-performance CPU and tend to have similar cooling requirements, while BGA-packaged chips (including GPUs and chipsets) have a tendency to pop away from a PCB if it flexes too much.

These. Laptops often use daughterboards for the GPU, to have a smaller board done on more layers. Alternately, they'll use a high layer count for the entire laptop mobo, and use the additional routing options to pack things in tighter.

Also, current high end GPUs output much more heat than CPUs, generally 1-10X as much. CPUs are also trending down in heat output while GPUs are holding steady, not decreasing.

I'd imagine cost is the major factor, especially from R&D. Anyone recall the recent(-ish) concept board Asus put out?
http://rog.asus.com/116882012/news/asus ... u-onboard/

TDIdriver wrote:

I'd imagine cost is the major factor, especially from R&D. Anyone recall the recent(-ish) concept board Asus put out?
http://rog.asus.com/116882012/news/asus ... u-onboard/

Ooooh, shiny!! I bet that thing costs a fortune, but it looks badass. It's a good example though. Super high costs, narrow market, no upgradeability....and I wonder how good the GPUs really are when they're passively cooled with relatively small heastinks compared to passively cooled discrete GPUs. I'm guessing they're mobile GPUs, and not very "high powered" at that. Why not make it liquid cooled?

Costs a fortune, looks badass, already out of date, no expansion slots whatsoever, it's like a laptop with no screen.

It very easily illustrates why no one does this except as a "Look at Me!" piece.

With two 8-pin and two 6-pin connectors, that Asus board could have the equivalent of a pair of GTX680's, Titans, or most likely, HD7970's. That's a lot of power for what looks like 1/10th the cooling of said above GPUs. I mean, there's 600W there between PCIe power connectors and PCIe slots, with less thermal dissipation area than a typical aftermarket CPU cooler designed for 150W.

Airmantharp wrote:

With two 8-pin and two 6-pin connectors, that Asus board could have the equivalent of a pair of GTX680's, Titans, or most likely, HD7970's. That's a lot of power for what looks like 1/10th the cooling of said above GPUs. I mean, there's 600W there between PCIe power connectors and PCIe slots, with less thermal dissipation area than a typical aftermarket CPU cooler designed for 150W.

Thus why it's a mockup, nonfunctional, and never sold to anyone?

As others have said, it's been done in the mobile sector extensively. You'd also find it back in the late '90s to very early '00s with some cheap consumer-grade boxes with stuff like ATI Rage Pros and the like (which also had their own discrete memory pools). But that ZEUS monstrosity notwithstanding, the only place you'll find it often in the desktop form factor any more is in some server motherboards that include some ultra-cheap, low TDP GPU design thrown in as an obligatory sop. If I remember correctly I saw a modern Opteron 'board with some kind of Matrox G200 chip on it just a few weeks ago...

Yes, G200MMS and ATI Rage Pro are common choices on server boards. They output a simple text console or VGA frame buffer very reliably, they draw a single watt, tops, need only 8MB or 16MB of vram to operate, and are documented to a ridiculous degree. They're also very simple designs, easy to fab even on low end machines, and very simple to diagnose/repair/replace.

Servers have a different set of criteria. By server criteria, that crap little G200 is superior to your GTX680 in so many ways that it is not even funny.

Forge wrote:

Yes, G200MMS and ATI Rage Pro are common choices on server boards. They output a simple text console or VGA frame buffer very reliably, they draw a single watt, tops, need only 8MB or 16MB of vram to operate, and are documented to a ridiculous degree. They're also very simple designs, easy to fab even on low end machines, and very simple to diagnose/repair/replace.

Servers have a different set of criteria. By server criteria, that crap little G200 is superior to your GTX680 in so many ways that it is not even funny.

Preaching to the converted here. As long as it outputs a 2D framebuffer and doesn't complicate the design, it's golden.

Although nowadays, the ATI "ES1000" is another popular server chip.

(For what it's worth, the "ES1000" is really a die-shrunken, PCIe-compatible RV100. As in ye olde Radeon 7000.)

And I've got a server coming in with an ASpeed AST2000, which I think is actually a new dumb 2D framebuffer design, with an ARM9 duct taped to it for IP management.