backtop

Hardware AMD Announces "Fusion" CPU/GPU Program
Anh Tuan Huynh (Blog) - October 25, 2006 4:46 AM
Print E-mail del.icio.us 69 comment(s) - last by mindless1.. on Nov 1 at 3:12 PM
The future of CPU/GPU computing

With the completion of AMD’s acquisition of ATI, AMD has announced its working on new CPU/GPU silicon that integrates the CPU and graphics processor into a single unit. The upcoming silicon is currently codenamed Fusion and is expected in the late 2008 or early 2009 time frame. AMD claims Fusion will bring:

AMD intends to design Fusion processors to provide step-function increases in performance-per-watt relative to today’s CPU-only architectures, and to provide the best customer experience in a world increasingly reliant upon 3D graphics, digital media and high-performance computing. With Fusion processors, AMD will continue to promote an open platform and encourage companies throughout the ecosystem to create innovative new co-processing solutions aimed at further optimizing specific workloads. AMD-powered Fusion platforms will continue to fully support high-end discrete graphics, physics accelerators, and other PCI Express-based solutions to meet the ever-increasing needs of the most demanding enthusiast end-users.

AMD expects to integrate Fusion for all its product categories including laptops, desktops, workstation, servers and consumer electronics products. Judging by the inclusion of PCI Express support, it would appear the integrated GPU is more of a value solution—similar to Intel’s cancelled Timna processor. It is unknown if AMD will retain the current Athlon and Opteron names with the launch of Fusion. This isn't too surprising as AMD and ATI previously promised unified product development including vague mentions of hybrid CPU and GPU products. AMD also previously announced its Torrenza open architecture as well.

In addition to Fusion, AMD expects to ship integrated platforms with ATI chipsets in 2007. The platforms are expected empower commercial clients, notebooks, gaming and media computing. AMD expects users will benefit from greater battery life on the next-generation Turion platforms and greater enhancements with AMD Live! systems. DailyTech previously reported on ATI's chipset roadmap which outlined various integrated graphics and enthusiast products.

With the development of Fusion and upcoming integrated AMD platforms, it is unknown what will happen to NVIDIA’s chipset business, which currently relies mainly on AMD chipset sales.

Comments     Threshold


This article is over a month old, voting and posting comments is disabled
good workstation solution?
By Kim Leo on 10/25/2006 5:40:57 AM , Rating: 3
we all probably knew this was comming, i look forward to it but i don't hope it becomes all they produce, since it would only make the cpu more expensive for people who don't want to use the "onboard" solution.. but maybe it'l be a decent(spell?) gfx solution, the awsome thing is that you no longer have to have some crappy PCI card around in case of you'r primary one is broken..




RE: good workstation solution?
By pinadski on 10/25/2006 6:03:56 AM , Rating: 2
hope this would be good for gaming graphics & physics at a less price
and hope not to use 1000w power supply


RE: good workstation solution?
By Grated on 10/25/2006 6:07:18 AM , Rating: 2
These will most probably be cores that can be found in integrated chipsets first...
Don't expect miracles :)


RE: good workstation solution?
By leidegre on 10/25/2006 6:18:08 AM , Rating: 2
Even if the GPU itself isn't targeting high-end performance ATI's physics move, with 3 GPU two for graphics and 1 for physics, this could prove to be used to accelerate performance in applications.

Additional software/game manufactures might need to consider opening up thier applications for greater customization, since platform are starting to grow in complextity and we are seeing basically on new solution for everything.


RE: good workstation solution?
By Kim Leo on 10/25/2006 6:25:31 AM , Rating: 4
the thing is that it would need to use system ram wich you probably already know, is much slower than those on high end graphics cards today.


RE: good workstation solution?
By otispunkmeyer on 10/25/2006 7:20:57 AM , Rating: 2
its no different to current IGP's in that respect, but with DDR2 and DDR3 on the horizon system bandwidth will be rather decent and combined with AMD's on board memory controller (i presume they will keep this) the latencies will be much improved

of course it wont hold a candle to a discrete solution, but i bet it'll be more than enough for casual gamers and most 3d CAD packages etc

PS PLEASE FIX THIS STIE!!!

should make the system more refined, less heat, less cost etc


RE: good workstation solution?
By Targon on 10/25/2006 7:31:40 AM , Rating: 2
By the time Fusion comes out, I expect that we will see DDR-3 or perhaps even a quad-bank memory controller on the CPU to help with this.

There are also two seperate workstation markets out there, the low end and high end. For low end workstations, then this will be a great solution that will cut costs. For the high end, if the CPU/GPU can work in Crossfire mode, then we may see the CPU/GPU plus TWO crossfire enabled graphics cards for a total of three GPUs.


RE: good workstation solution?
By Nocturnal on 10/25/06, Rating: 0
RE: good workstation solution?
By Spivonious on 10/26/2006 4:50:59 PM , Rating: 4
I don't know because that would take them right out of the enthusiast market. Why would I upgrade my CPU everytime I wanted a new graphics card?


RE: good workstation solution?
By jp7189 on 10/25/2006 10:52:02 AM , Rating: 2
quote:
the thing is that it would need to use system ram wich you probably already know, is much slower than those on high end graphics cards today.


The path currently is: CPU -> chipset -> PCIe -> GPU

Having direct CPU/GPU communications will vastly improve any app that doesn't use large testures - the vast majority of users.


RE: good workstation solution?
By ogreslayer on 10/25/2006 2:57:24 PM , Rating: 2
You are forgetting that unless both have access to memory through the IMC; you still have at least one of those steps. No overhead would actually be removed as you are now flooding the HT interconnect. To top that off even if they go Quad-Channel by the time they start integrating, the bandwidth is not even close to what R600 is supposed to have. DDR3 and 4 are not gonna be any kind of salvation as all these offer only efficency, they are still DDR and all we will get is a speed bump. We can pray for lower latencies.

AMD is not a big enough player to be the one who starts the move to the next flavor of RAM. Intel has to do it and increasing the memory speed has proven to have marginal effects on Core 2 products when compare to Athlon64X2s. I wouldn't expect any real rush from Intel until we get the 1333FSB and they move to a 256-bit bus. This is gonna be integrated graphics level stuff. Unless Nvidia can pump out a CPU; Intel and AMD/ATI are gonna smother them out of the integrated market for desktops and notebooks.


RE: good workstation solution?
By Viditor on 10/29/2006 7:21:54 PM , Rating: 2
quote:
unless both have access to memory through the IMC; you still have at least one of those steps. No overhead would actually be removed as you are now flooding the HT interconnect

I don't think people yet understand the way Fusion will work in tandem with Torrenza yet...
Phil Hester explained a bit to TheReg in this interview:
http://www.reghardware.co.uk/2006/10/26/the_story_...
Notice that Fusion will be only a part of the modular design...
Very good article...


RE: good workstation solution?
By Rollomite on 10/25/2006 10:07:02 AM , Rating: 1
quote:
AMD intends to design Fusion processors to provide step-function increases in performance-per-watt relative to today’s CPU-only architectures


It would appear that they are going to try to keep power usage at the same level while incorporating this new technology.

Rollo


Vertex Shaders
By dunno99 on 10/25/2006 11:30:24 AM , Rating: 4
I think more credit should be given to this setup. This solution lends to the possibility of breaking up GPU processes instead of merging them. As in, in addition to taking a small chunk of unified shaders and put them on the CPU, and be able to output directly from that (which was given), AMD could also segregate the on-chip shaders to purely vertex processing and relegate the geometry and fragment shading to the graphics card. Due to the nature of the GPU, data feeding is pretty much one way from the vertex shader to the rasterizer and geometry/fragment shaders. This means that the CPU to GPU overhead could be drastically reduced by processing vertex instructions on the CPU/GPU hybrid, then sending the data down through the PCIe bus to the add-on graphics card, cutting driver processing time by even half, I would say (vertex data shouldn't take that much bandwidth, I don't think).




RE: Vertex Shaders
By wired009 on 10/25/2006 12:14:25 PM , Rating: 2
The Fusion solution seems convenient and beneficial at first glance but you have to wonder if it's practical to implement. Imagine a batch of new processors. Say I want high CPU performance but don't need superior graphics because I don't play games. Will there be fast CPU - low end GPU, fast CPU - mid range GPU, fast CPU - fast GPU variations so mainstream users and gamers have a choice? What happens during the next CPU refresh and the fast CPU is now the low end or midrange CPU? It will be hard for AMD to contine to offer a certain variation if it is no longer in high demand. This is where Fusion begins to look like a very cost ineffective solution for AMD. It makes a lot more sense to keep CPU and GPU separate for marketing reasons and to keep manufacturing lines efficient. It is more likely that computers will move towards removable socket GPUs that attach directly to the motherboard with the elimination of AGP/PCI slots than they are to move towards Fusion.


RE: Vertex Shaders
By NullSubroutine on 10/27/2006 7:43:00 AM , Rating: 1
it actually makes sense if you consider if u need more graphical horsepower you could slap in an amd made (or nvidia) gpu in the torenzza's 'accelerator' socket.


RE: Vertex Shaders
By sdsdv10 on 10/25/2006 12:43:58 PM , Rating: 2
quote:
This means that the CPU to GPU overhead could be drastically reduced by processing vertex instructions on the CPU/GPU hybrid, then sending the data down through the PCIe bus to the add-on graphics card,


Wouldn't incorporating a CPU/GPU hybrid mostly eliminate the need for an external graphics card? Otherwise, this will only be a high end product. The added cost of a CPU/GPU hybrid processor and extra graphics card would really raise the price on an overall system. Maybe I misunderstood, I thought they were going after the integrated graphics market (lower end) with a more elegant and efficient solution. But then again, I'm not really a "computer" guy.


RE: Vertex Shaders
By Tlogic on 10/28/2006 7:54:45 AM , Rating: 2
"Wouldn't incorporating a CPU/GPU hybrid mostly eliminate the need for an external graphics card?"

Yes and they already exist in integrated chipsets. The problem is can you shut off the 'gpu' on the cpu if you want to add a stand alone graphics solution?

Lastly stand alone graphics solutions will always be superior simply because you cannot get dedicated ram and high bandwidth bus's on a cpu/gpu that is integrated, main memory speed and/or bandwidth would have to increase in exteme amounts to catch up to stand alone solutions.


RE: Vertex Shaders
By MarcLeFou on 10/25/2006 1:56:33 PM , Rating: 3
Actually what I find interesting about this concept is that you can have a basic GPU core integrated into the CPU core which would be sufficient for everyday business applications, for basic workstations, for business laptops and for barebones computers which should cut costs for over 75% of all systems sold.

But what I find really smart about this concept is that, with the Torenza initiative, the CPU will now be able to communicate directly through the Hypertransport link with a bunch of addon cards. Most people so far have envisonned putting in a second or third GPU but what I see happening is actually a breaking down of the components of the GPU into separate parts. Apart from the obvious idea of increasing VRAM through an add-on card, think about being able to customize your GPU according to you usage scenario with specialized shader cards, geometry cards, mhz boosts, etc.

This system would be the ultimate in customization and would be much more price efficient for customers who would be able to get exactly what they need. And instead of changing a whole GPU when a new tech comes out, you could just change that particular add-on card giving a much longer lifespan to your video card, hence your system. Imagine just being able to upgrade to shader model 5,0 (or whatever it is then) just by changing your 50$ shader card instead of your whole video card like we have to today!

Also, assuming the technical hurdles can be overcomed, AMD would be the only one for a few cycles with this tech, creating a totally new market a bit like Nintendo is trying to do with its Wii and taking total control of it by catching the competition off guard because it would take Intel at least a year to develop a competing product in the best case scenario. Disruption of an established market to gain the leadership in both CPU architecture and GPU add-on cards in one fell swoop. Quite a business strategy.


RE: Vertex Shaders
By tbtkorg on 10/25/2006 4:24:23 PM , Rating: 5
Interesting thread.

Integrating the GPU with the CPU is not all about graphics; it's about making the tremendous parallel processing power of the GPU available for general computation, including graphics. Admittedly, I cannot imagine all the different applications for such parallelism any more than you can. Scientific computation will use it, at least, but it goes far beyond that. The belief is that the general-purpose GPU is inherently, fundamentally such a sound concept that people like you and me will soon come up with a thousand creative ways to put it to work, given the chance.

Readers who have written assembly code or programmed microcontrollers will best understand the point I am trying to make, because at the lowest programming level, GPU programming differs radically from traditional CPU programming. The CPU is code-oriented; the GPU, data-oriented. Wherever the quantity of data is large and the parallel transformation to be applied en-masse to the data is relatively simple, the general-purpose GPU can, at least in theory, greatly outperform any traditional CPU. The CPU, of course, is far more flexible, and still offers by far the best way to chain sequential calculations together. The marriage of the CPU to a general-purpose GPU is thus a profound concept, indeed.

The general-purpose GPU is an idea whose time has come. By acquiring ATI, AMD makes a serious attempt to dominate the coming generation of computer technology, taking over Intel's accustomed role as pacesetter and standard bearer. Of course there is no reason to expect Intel to sleep through this transition. If Intel responds competently, as one assumes that it will, then we are in for some very interesting times in the coming few years, I think.

There is a third element, besides the CPU and the GPU, which will emerge soon to complement both, I think. This is the FPGA or field-programmable gate array. Close on the heels of the CPU-GPU marriage, the integration of the FPGA will make it a triumvirate, opening further capabilities to the user at modest additional cost.
AMD/ATI will not be able to ignore this development, even if their general-purpose GPU initiative succeeds, as I think it will. Interesting times are coming, indeed.