The Outlook On AMD's Fusion Plans

PreacherTom writes "Now that AMD's acquisition of ATI is complete, what do the cards hold for the parent company? According to most experts, it's a promising outlook for AMD . One of the brightest stars in AMD's future could be the Fusion program, which will 'fuse' AMD's CPUs with ATI's GPUs (graphics processing units) in a single, unified processor. The product is expected to debut in late 2007 or early 2008. Fusion brings a hopes of energy efficiency, with the CPU and GPU residing on a single chip. Fusion chips could also ease the impact on users who plan to use Windows Vista with Aero, an advanced interface that will only run on computers that can handle a heavy graphics load. Lastly, the tight architecture provided by Fusion could lead to a new set of small, compelling devices that can handle rich media."

Stock tip ...

Invest in heat sinks! :-)

Re:Stock tip ...

Yeah, but the heatsink for the processor/graphics card combo system will be righteous.

Frankly, I'm betting this is going to turn out more like the next generation of integrated video. Basically, the only "fusion" chips you'll see will be ones designed for small/cheap boxes that people never upgrade the components on. I'm betting the graphics in general will be slow and the processor will be average. Super fast processors and fast graphics won't get the fusion treatment because the people who buy them tend to want to keep them separate (for upgrading later), not to mention the difficulty you'd have powering and cooling a chip that complex.

Airport fun

how will homeland security like you bringing home a multi core Fusion through the gates?

"But, but its an AMD processor, built in Germany or Russia or somewhere"

"Teh internet told me it was more powerful than anything else out there."

"It would literally blow me away!"

Re:Airport fun

You assume that this would do away with video cards; there's not a chance of that happening any time soon. As I said in another thread, it'd be quite simple for AMD to disable the on-chip video in favor of a detected add-in card.

Right now I'm buying a $200 vidcard every 18-24 months. I'm looking at probably getting my next one middle of next year, around the same time I replace my motherboard, CPU, and RAM. My current system is struggling with the Supreme Commander beta and upcoming games like Crysis should be equally taxing on it. In the past six years, I've bought three CPU upgrades. If AMD could market a $300 chip that gave me a CPU and GPU upgrade with similar performance and stay on the same socket for 3-4 years, I'd be breaking even.

A better way to spend they money would be on PR...

.. or advertising on TV. I work in a computer shop and it seems loads of people have no idea who the hell AMD are. I've explained that they're just competitors to a lot of customers, but still the customers go 'No, I've been told to get an Intel.' I can't recall having ever seen an AMD ad on telly at all.

power efficiency??

yeah. i'm also wondering how putting the two hottest components on the mother board (the GPU and CPU) into the same package is a power savings... :-/ maybe on the low end of the market where the performance of the GPU is irrelevant, but for those who actually care about GPU performance, putting the two most power hungry and memory bw hungry components together doesn't seem like a good idea.

Re:power efficiency??

Any kind of integration tends to improve power efficiency just because of the high capacitance of the PCB traces. This makes it difficult to route a PCB for high-speed inter-chip communications never mind getting multiple 2.5Gb/s (PCIe) signal traces through a connector. All this requires large driver cells to drive off-chip communication and these use a great deal of power (and moderate area) on chip. Reducing the noise floor of your signals (by keeping them on chip) also gives you more headroom for voltage reductions in your digital hardware. All in all it makes it a much better picture overall for power efficiency. But dissipating power from these new chips will still be a headache for CPU package designers and systems guys alike.

Bad idea for upgrades

Although I can see the potential efficiency increases, combining the GPU and CPU into one chip means that you will be forced to upgrade one when you only want to upgrade the other. To me, this seems like a bad idea in that AMD would have to make dozens of GPU/CPU combinations. Say I want one of AMD's chips in my headless server, am I going to have to buy a more expensive processor because it has a high powered GPU that I don't want or need? What if I want to build a system with a good processor to start, but due to budget reasons want to hold off on buying a good video card?

Combining the CPU and GPU may make sense for embedded systems or as a replacement for integrated graphics, but I cannot see it working for those who prefer to have specific components based on other factors.

It's for laptops and budget systems

Especially the former, where you can't really upgrade anyway and you typically have a GPU soldered to the board.

The advantages of a combined CPU/GPU in this space are:
1) Fewer chips means a cheaper board.
2) The GPU is connected directly to the memory interface, so UMA solutions will not suck nearly as hard.
3) No HT hop to get to the GPU, so power is saved on the interface and CPU-GPU communication will be very low latency.

I highly doubt AMD is planning on using combined CPU/GPU solutions on their mainstream desktop parts, and they are absolutely not going to do so for server parts. I think in those spaces they'd much rather have four cores on the CPU, and let you slap in the latest-greatest (ATI I'm sure they hope, but if NVidia gives them the best benchmark score vs Intel chips then so be it) graphics card.

AMD has already distinguished their server, mobile, desktop, and value lines. They are not going to suddenly become retarded and forget that these markets have different needs and force an ATI GPU on all of them.

Re:It's for laptops and budget systems

I highly doubt AMD is planning on using combined CPU/GPU solutions on their mainstream desktop parts, and they are absolutely not going to do so for server parts

I think they are, and I think it's the right choice. The GPU that will be integrated will not be today's GPU, but a much more general processor. Look at NVidia's G80 for the beginning of this trend; they're adding non-graphics-oriented features like integer math, bitwise operations, and soon double-precision floating point. G80 has 128 (!) fully general-purpose SISD (not SIMD) cores, and soon with their CUDA API you will be able to run C code on them directly instead of hacking it up through DirectX or OpenGL.

AMD's Fusion will likely look a lot more like a Cell processor than, say, Opteron + X1900 on the same die. ATI is very serious about doing more than graphics: look at their CTM initiative (now in closed beta); they are doing the previously unthinkable and publishing the *machine language* for their shader engines! They want businesses to adopt this in a big way. And it makes a lot of sense: with a GPU this close to the CPU, you can start accelerating tons of things, from scientific calculations to SQL queries. Basically *anything* that is parallelizable can benefit.

I see this as nothing less than the future of desktop processors. One or two x86 cores for legacy code, and literally hundreds of simpler cores for sheer calculation power. Forget about games, this is much bigger than that. These chips will do things that are simply impossible for today's processors. AMD and Intel should both be jumping to implement this new paradigm, because it sets the stage for a whole new round of increasing performance and hardware upgrades. The next few years will be an exciting time for the processor business.

So...

Energy efficiency...
Project named Fusion...
...
Please tell me Pons and Fleischmann [wikipedia.org] aren't behind this?

Heat???

Although CPUs have gotten better in the past year, GPUs (particularly ATI's) still keep outdoing each other in just how much power they can suck.

With a decent single-GPU gaming rig drawing over 200W just between the CPU and GPU, do they plan to start selling water cooling kits as the stock boxed cooler?

Re:Heat???

Although CPUs have gotten better in the past year, GPUs (particularly ATI's) still keep outdoing each other in just how much power they can suck.

You're talking about the high-end "do everything you can" GPUs... ATI is dominating the (discrete) mobile GPU industry because their mobile GPUs use so little power. Integrating (well) one of those into a CPU should still result in a low-power chip.

Yes but

Will it run Linux less than half a year after it's obsoleted by the next version?

Upgrades ?

Let's say I buy Fusion. Later on NVIDIA brings cool graphics card to market. Will I be able to use NVIDIA graphic card with Fusion ?

Disaster for Linux and OSS

Whoa. You're going to need a closed-source kernel driver to use your CPU now? They can eat me. The graphics driver situation is bad enough.

This one is untouchable until they open up the graphics drivers - or goodbye AMD/ATI.

jh

but...

does this mean ATI will be opening up its GPU programming specs, or merely what is being stated (that graphics chip and CPU will share a die) ?

Remember math coprocessors?

Prior to 486s, they used to have the floating point functions on a separate chip from the processor. If the GPU is moving to the processor now, what will be the next thing to get sucked in?

cool

this processor will just be a addon ontop of the vid card you would have, like 3dnow! its just a add on, maybe we can shift into 128bit processors!

Maybe...

The article says that this might be attractive to businesses: I can see that since most businesses don't care about graphics. This is similar to businesses buying computers with cheap on-board video cards. But that means they will be profiting on the low-end. It seems like this is more of a boon for laptops and consoles: Currently, laptops with decent video cards are expensive and power-hungry. Same with consoles. But for mid-range and high-end systems, there must be a modular bus connecting these two parts since they are likely to evolve at different rates, and likely to be swapped-out individually.

Linux Drivers

I've been an nVidia advocate since 1999 when I bought a TNT2 Ultra for playing Quake III Arena under Linux on my (then) K6-2 400.

I'm on my 4th nVidia graphics card, and I have 6 machines, all running Linux. One is a 10-year-old UltraSPARC, one has an ATI card.

Despite slashbot rantings about the closed-source nVidia drivers, and despite my motley collection of Frankenstein hardware, I've never had a problem with the nVidia stuff. The ATI stuff is junk. The drivers are pathetic (open source) and the display is snowy, and the performance it rubbish.

I hope AMD do something about the Linux driver situation.

My next machine will be another AMD, this time with dual dual-core processors and I'll be doing my own Slackware port, but I'll be buying an nVidia graphics card.

Cyrix MediaGX

Cool amd is about to MediaGX themselves.

this will fail

the processor market still changes too rapidly for this kind of bonding.

Do you really want to have to replace an entire system when you upgrade? You buy a Dell, a new game comes out 6 months and your system can't play it reasonably well.

So then you either
a) buy a new system
or
b) gut in a video card and not use the one on the proc.

When processors begins to peak, and each upgrades is basically a few ticks, then developers will have to create things for the systems that is out , not a system that will be out in a year.
When this happens*(and it will) software will enter a golden age.

Of course, someone could come up with completly different technolgy and make current procs. irrelevant..

*
There are many factors coming to light that are already slowing down proc development. Die limitations, noise limitations, bus size limitation, to just name a few. Don't confuse practicallity with theory.

At the risk of being modded reundant

I'm going to ask:

That's great and all, but does it run Linux?

I'm not kidding, either. Is AMD going to force ATI to open up its specs and its drivers so that we can FINALLY get stable and FULLY functional drivers for Linux, or are they still going to be partially-implemented limited-function binary blobs where support for older-yet-still-in-distribution-channels products will be phased out in order to "encourage" (read: force) customers to upgrade to new hardware, discarding still-current computers?

That is why I do not buy ATI products any more. They provide ZERO VIVO support in Linux, They phase out chip support in drivers even while they are actively distributed. They do not maintain compatibility of older drivers to ensure they can be linked to the latest kernels.

This is why I went Core 2 Duo for my new system and do not run AMD - their merger with ATI. My fear is that if ATI rubs off on AMD then support for AMD processors and chipsets will only get worse, not better.

GPU or GPGPU?

From what I understand (and I could be wrong), AMD/ATI is aiming more at the GPGPU market. So we're talking more of a suped up altivec processor in the CPU instead of a full blown GPU. It sounds like the're simply adding a 64 pipleline vector processor to the existing x86-64 core. I'm not sure if this is a bad idea.

I remember programming assembly graphics code in BASIC back in the day. You would set the VGA card to mode 13h and then write to...what was it now...0xa00? That's probably wrong. Anyway, whatever your wrote to that portion of memory would go to the screen.

If you had a huge SIMD co-processor, would it not be possible to rival modern GPUs with this model? Not to mention being able to do some cool stuff like having a video input card dump data driectly into that portion of the screen. So you could have video in with the CPU at complete idle.

Not what you think

My people are reading this as an integrated GPU and CPU. I don't see it that way. I see it as adding a generic vector processor to the CPU. Similar to the Cell processor and similar to future plans Intel has described. Vector processors are similar to SSE, 3DNow, etc. They are SIMD processors that can execute very regular mathematical computations (Video and audio encoding/decoding) VERY quickly, but aren't much good for generic algorithms.

A step between on-board video, and full graphics

The people claiming this will fail all seem to miss the market this is aimed at. It's obviously not intended to compete with the high-end, or even middle of the road graphics processor. Those boards require gobs of VERY fast video memory. My guess is this thing is aimed at a space between the on-board video (which are really just 2-d chips) and the full 3-d graphics card. Anyone buying this has no intention of buying a super-duper

With Vista coming out soon, PC-makers are going to want a low-cost 3-d accelerated solution to be able to run some (or maybe all) of the eye-candy that comes with vista.

I'll buy it if they provide free drivers

I'll buy this is they provide free drivers; I won't buy it if they don't. Vista's piggish graphics will surely push all GPU's to new performance levels. I don't care about on-chip integration nearly as much as I care about avoiding the need to use binary blobs in my free OS.

For the sake of competition...

Let's hope this fusion doesn't bomb.

How many blades?

AMD will be making razors and shave gel? Sweet! How many blades, 4, 5 or scalable on demand?

Fusion or Design By Committee (aka "Convergence")

Great idea on paper. It boils down to personnel though. You're talking about fusing development teams with experience. Will they work together well? Or will the elevator assets go work for someone else, leaving the understudies to bicker about with an ignoramus boss unable to figure out which engineers are clever and which are just suckups?

I'm not saying it won't work; I'm saying that fusing development teams with expertise is a lot different than fusing different components onto the same board. And that, in turn, is a lot different than a multi-option fuze.

Dual and Quad socket!

I know the trend is single socket multi core but with the gpu embedded dual and quad sockets instead of sli!

My fear is...

This will lead to a whole new world of disgustingly bad graphics chips eating system RAM and claiming to have "256MB" or whatever but really having little or none and just munching on (slow) system memory as needed... and that never works as well as it should.

Matrix Operations?

How about getting those lightning fast matrix operations onto the CPU? I always hear about people building application-specific tweaks by reprogramming their algorithm into a shader language. I imagine that there is far more fertile soil for innovation here than some lame combined C/GPU.

VIDEO cards on the HT bus

They are also look at this as well. Maybe even some kind of an super cross fire / sli with build in cpu video graphics processing + video in the slot with it's own graphics processing and ram also you may be able to 2 video cards linked as well.

Amd 4x4 systems may be able to have 4 video cards + 2 cpus with graphics processing in them.

Is this when Apple moves to AMD?

Interestingly there was an interesting article on MacRumors about one of Apples suppliers ordering more capacitors that were associated with AMD chips. Although the source is noted to be a bit unpredictable.

http://www.macrumors.com/pages/2006/11/20061116133 114.shtml [macrumors.com]

I would think such a design would be a great chip for laptops, minis, iMacs and iTV!