Graphics-Enabled CPUs To Take Off In 2011

angry tapir writes "Half the notebook computers and a growing number of desktops shipped in 2011 will run on graphics-enabled microprocessors as designers Intel and Advanced Micro Devices (AMD) increase competition for the units that raise multimedia speeds without add-ons. The processors with built-in graphics capabilities will be installed this year on 115 million notebooks, half of total shipments, and 63 million desktop PCs, or 45 percent of the total, according to analysts."

Comments

Supercomputing

[louic]

Depending on how exactly these processors will look like, they may be very interesting for speeding up scientific computations. The fastest computer in the world at this moment is already GPU based, and such a CPU/GPU hybrid can possibly be even more efficient by removing the slow communication between CPU and GPU.

Overheating already...

[Anonymous Coward]

...CPU handling the graphics in laptops is already causing overheating issues.

Two cases in point, a Toshiba laptop with AMD and a 13" MacBook Pro with Intel, the fans run annoyingly at high speed, the bottoms are hot enough to fry eggs on. That's just sitting with one web page open. How long can one expect machine like that to last? A year? two maybe?

Are web pages going to suddenly tone down their act, quit using video, animation, Flash? Text and pictures only? If they do that, then what? Hardware makers only start making laptops that can handle web text?

Dedicated graphics is the way to go, CPU and graphics on separate dies away from each other, separate the heat sources.

I can just imagine the scene where a bunch of power hungry types just made the decision to move towards integrated graphics, and a highly intelligent engineer just stomping out of the boardroom in protest.

The more things change...

[Anonymous Coward]

Way back near the dawn of time, Intel created the 8086, and its slightly less capable little brother, the 8088. And they were reasonable processors ... but although they were good at arithmetic, it was within tight constraints. Fractions were just too hard. Trigonometry sent the poor little souls into a spin. And so on.

And thus, the 8087 was born. It was able to carry the burden of floating point mathematical functions, thereby making things nice and fast for those few who were willing to pony up the cash for the chip.

Then out came the 80286 (let's forget about the 80186, it's not really all that relevant here). It was better at arithmetic than the 8086, but still couldn't handle floating point - so it had a friend, the 80287, that filled the same purpose for the 80286 as the 8087 did for the 8086 and 8088. (We'll blithely ignore Weitek's offerings here. They existed. They're not really germane to the discussion.)

Then the 80386. Much, much better at arithmetic than the 80286, but floating point was still an Achilles heel - so the 80387 came along for the ride.

And finally, the i486. By this stage, transistors had become small enough that Intel could integrate the FPU on die - so there was no i487. At least, not until they came out with the i486SX, which I'll blithely ignore. And so, an accelerator chip that was once hideously expensive and used only by a few who really needed it was integrated onto chips that everybody would buy.

Funnily enough, it was around the time that the i486 appeared that graphics accelerators came onto the scene - first for 2D (who remembers the Tseng Labs W32p?), and then for 3D. Expensive, used only by a few who could justify the cost ... is this starting to sound familiar to you?

So another cycle is beginning to complete, and more functionality that used to be discrete is now to be folded onto the CPU. I can't help but wonder ... what will be next?

Re:A GPU by any other name would render as slowly

[Luckyo]

I actually can chip on this on a "this is not true" side. My father isn't a gamer by any stretch - the only games he likes to play are various arcanoid derivatives. Which meant that his work laptop served him just fine.
Then came shatter, and he all but killed me with his "why won't my laptop run this?" questions. Try to explain to someone running the crappy intel 945GM that always ran the old 2d arcanoids that shatter just won't work on it.

So now, I'm probably giving them my current gaming computer as I upgrade, and I'm pretty sure he'll be telling tech support at work that his next laptop has better include 3d acceleration or else (he's in position to be able to tell them that). So the old saying applies here - you'll be satisfied with integrated, until in comes one killer application that it won't run, and then you aren't. Problem is, with so much software requiring decent 3d graphics on board (even aero does!) you're still best served by a half decent dedicated graphics card that powers itself down when 3d features aren't used or used sparingly.

Finally there's an issue of quality, and that goes beyond 3d. Most integrated chipsets have clear problems displaying higher resolutions, which is why high resolution laptops generally have a dedicated chipset rather then integrated solution.

Re:A GPU by any other name would render as slowly

[giuseppemag]

First of all where would this be FUD? Try connecting a full HD monitor to an integrated Intel GPU and you'll see what I meant.

Also, this bullshit that users don't do computing intense stuff is, well, bullshit. Full HD video, 3D movies, photo processing are computationally intensive even if they are not particularly serious usage of computing power. Don't confuse "important work" with "computationally intensive work".

Re:A GPU by any other name would render as slowly

[Auroch]

My back-of-the envelope calculations tell me that the 9W version is at least as powerful as a low-end Nvidia 400-series or ATI 5000-series

My back of the envelope memory tells me that all low end 400 series and low end 5000 series "graphics" are actually IGPs as well...

Re:And the advantage is...?

[petermgreen]

And the advantage is...?

The advantage of shared memory graphics is reduced cost and power consumption.
The advantage of integrating the memory controller in the CPU is it allows the CPU faster access to memory.
The advantage of reducing the number of high speed chips is reduced cost and power consumption.

So with that in mind lets consider the options for a CPU with an integrated memory controller.

Putting the shared memory graphics on a seperate chip would require a link to the CPU that offered high speed high priority ram access by the GPU and would still leave you with two high speed chips. AMD do this with hypertransport though IIRC they usually have a small ammount of dedicated graphics memory as well to keep the framebuffer traffic off the hypertransport links.

Not offering shared memory graphics at all rules a platform out of the low end market and makes it less than ideal for the business market in general. Intel did this with the nahelm quad and hex core processors and I belive are planning to do the same with the LGA2011 high end sandy bridge chips.

So the natural thing to do is to put the shared memory graphics on the CPU with the memory controller. Intel did this with the dual core nahelm chips and with the LGA1155 mainstream sandy bridge chips.

So there will be more computers with crappy integrated graphics.

Probablly a few more because there were no nahelm quad cores with integrated graphics support. So if you wanted a fast quad core you pretty much had to have discrete graphics as well whether you wanted them or not.

Practically speaking sandy bridge puts things pretty much back the way they were before with the choice of processor core count decoupled from whether to use integrated graphics. It's just those integrated graphics are in the CPU rather than the northbridge. Hopefully this will mean the likes of dell will finally migrate off LGA775.

Oh, and btw, wasn't the plan until recently to basically replace the CPU with the GPU?

GPUs are great at some types of calculation but suck at branch heavy code. So many algorithms have to be completely redesigned to run on them. IIRC in the case of video encoding GPUs can do it quicker but only using cut down encoders that produce lower quality results.

AMD was at one point planning to make units that combined the best of both (note: the fusion name which originally reffered to this is now being used to reffer to CPUs and GPUs on the same die but logcially seperate). Dunno if they still are.