Slashdot Mirror
AMD's Fusion Processor Combines CPU and GPU
ElectricSteve writes "At Computex 2010 AMD gave the first public demonstration of its Fusion processor, which combines the Central Processing Unit (CPU) and Graphics Processing Unit (GPU) on a single chip. The AMD Fusion family of Accelerated Processing Units not only adds another acronym to the computer lexicon, but ushers is what AMD says is a significant shift in processor architecture and capabilities. Many of the improvements stem from eliminating the chip-to-chip linkage that adds latency to memory operations and consumes power — moving electrons across a chip takes less energy than moving these same electrons between two chips. The co-location of all key elements on one chip also allows a holistic approach to power management of the APU. Various parts of the chip can be powered up or down depending on workloads."
Heh. They should use Apu from the Simpsons in their advertising...
“Hundreds of millions of us now create, interact with, and share intensely visual digital content,” said Rick Bergman, senior vice president and general manager, AMD Product Group. “This explosion in multimedia requires new applications and new ways to manage and manipulate data."
So people watch video and play video games, and it's still kinda pokey at times. We're way past diminishing marginal returns on improving graphical interfaces.
I bring it up, because if you're trying to promote a technology that actually uses a computer to compute, you know, work with actual data, you are perpetually sidetracked by trying to make it look pretty to get any attention.
Case in point: working on a project to track trends over financial data, there were several contractors competing. One had this software that tried to glom everything into a node and vector graph, which looked really pretty, but didn't actually do anything to analyze the data.
But to managers, all they see is that those guys have pretty graphs in their demos and all we had was our research into the actual data... all those boring details.
"Moving electrons between two chips" isn't entirely accurate. What moves is a wave of electric potential; the electrons themselves don't actually move very far.
I'm hoping moving things into the CPU will make it easier to take advantage of the huge parallel architecture of modern GPUs.
For what, you ask?
I'm personally interested in sound synthesis. I play the piano, and while you can get huge sample libraries (> 10 GB), they're not realistic enough when it comes to the dynamics.
Instead people have been researching physical models of the piano. So you simulate a piano in software, or the main components of it, and extract the sound from that. Nowadays there are even commercial offerings, like Pianoteq (www.pianoteq.com) and Roland's V-Piano. Problem is that while this improves dynamics dramatically, they're not accurate enough yet to produce a fully convincing tone.
I think that's partly because nobody understands how to model the piano fully yet, at least judging from the research literature I've read, but also very much because even a modern CPU simply can't deliver enough FLOPS.
Perhaps you should email your insights the CEO of AMD. I'm sure he'll be grateful for the heads up from some retarded cunt on slashdot that his huge array of engineers and scientists have being building a chip that doesn't work for the past 5 years.
It doesn't bring anything to the table yet. Firstly, IOMMUs need to be more prevalent in hardware, then secondly there needs to be support for using them in your favourite flavour (Xen will be there first) of virtualisation.
That said, we'll get ugly vendor-dependent software wrapping of GPU resources. Under the guise of better sharing of GPUs between VMs, but really so you're locked in.
The technical difference is that while your Core i3 has its GPU as a separate die in the same packaging, AMD Fusion has the GPU(s) on the same die as the CPU(s). The Intel approach makes for shorter and faster interconnects, the AMD approach completely removes the interconnects. The main advantage is probably (as is alluded to in the summary) related to power consumption.
...ceterum censeo Carthaginem esse delendam.
Let's party like it's 1995! Again!
Slightly less cynically, isn't this (in like-for-like terms) trading a general purpose CPU core for a specialised GPU one? It's not like we'll get more bang for our buck, we'll just get more floating point bangs, and fewer integer ones.
If you were blocking sigs, you wouldn't have to read this.
I also heard that they *share* FP math units between CPU and GPU.
AMDs product is just a desperate attempt at trying to be relevant. They need to show they have a product competing with the big boys in all the right channels.
AMD is plenty relevant. It is Intel that scrambled to put out a 6 core desktop processor, which was so poorly planned that the cheap version is $1000. Meanwhile nVidia is desperately trying to get people locked into their CUDA API because their video cards just dont bang the performance drum like they used to.
AMD and Intel have different visions. AMD is clearly focusing on getting more cores on chip for more raw parallel performance (12 core CPU's in 4 chip configs are owning the top end server market.. brought to you by AMD), while Intel is clearly trying to maximize memory bandwidth to peak out raw single threaded performance (triple channel ram and larger cache is owning the software rendering and gaming markets)
Normal people are within the $50 to $200 CPU range, and at those price points, solutions from both camps perform about the same. On the video card front, you just can't beat AMD right now. Best price/performance ratio on top of best performance period.
"His name was James Damore."
How so?
AMD's offer is real, it uses a real performant GPU, not a GMA joke. Larrabee is stil vapourware, and it will be for a long time.
The 6-core Intel processor is the Extreme Edition (always was introduced at $1000), and frankly smokes every other desktop processor out there.
AMD is the value-choice - they're cheaper at the same performance point, but they don't really compete in the over $250 desktop arena.
On the server front, Intel's introduction of Core2 based Xeons allowed it to compete again, and right now AMD is leader only in some cases in server performance (some are draws, but most I think go to Intel). Too bad, as server processors were producing a lot of money for AMD.
Intel is also leader in performance/watt, due to a complex power delivery architecture and better processor production facilities.
Meanwhile, AMD competes where it can on the processor front (but ruled the previous 6 months on the performance graphic front).
Will the drivers for the graphics be open source or will we be crawling out of this proprietary driver hole we have been trying to climb out of for over a decade?
Indeed.
This GPU-on-the-CPU is targeting the mobile/lightweight market.
Think about how the other solutions work. That GPU chip sits next to the CPU chip and they both must be connected to the system bus in order to access ram. With AMD's solution here, you remove that GPU chip and therefor also remove the external BUS connection that it required. This is a very big win for manufacturers, who would even pay a premium for the chip because of the lower production costs. But knowing AMD, they wont be charging a premium for it. Instead they will try to push Atom's out of the market.
"His name was James Damore."
AMD chipsets with integrated GFX were quite good at power consumption already; using a dozen or so watts. Considering AMD puts out quadcores with sub 100W TDP, Fusion shouldn't be that big a problem.
One that hath name thou can not otter
Sure you can offload GPU work so long as the entire process is handled by the server and just streaming the result to the client. I've seen this done over the Internet besides on a LAN. It'd be different if it were trying to use the client CPU and memory to drive the GPU.
They specifically were pointing out the benefit to having the GPU and CPU on the same chip which is quite a bit different than a mobo integrated solution. It probably isn't as powerful as a Xeon quad-core processor and a $500 video card but the question is how well it is setup to handle many different GPU tasks. I'd at least assume it's quite a bit faster for these types of tasks than a standard CPU and I wonder how well they can scale the technology for a better CPU and GPU.
I'm not sure I agree it's a niche market. I'd say more of a market poised to explode when the right products make it attainable. For virtualization it's more important that it can handle several unrelated tasks at a reasonable speed than that it can handle a single task at a high speed. If each CPU core also had a paired GPU it'd open up possibilities. Bulk, power consumption, and heat are often as big of issues for server farms as for laptops which is another reason why an interpreted GPU might be of interest.
Grid computing uses goes hand in hand with virtualization. Again coming down to how well these can work in parallel. Being able to fit a number of CPU and GPU cores on a single physical chip could be very beneficial I think.
At what price learning? At what cost wisdom? The price is a man's peace of mind, and the cost is his life.
Well, "incorporating a better GPU" makes quite a bit of difference, considering i3/i5 solution isn't much of an improvement almost anywhere (speed - not really, cost - yeah, I can see Intel willingly passing the savings...anyway, cpu + mobo combo hasn't got cheaper at all, power consumption is one but mostly due to how Intel chipsets were not great at this); and seemed to be almost a fast "first" solution, announced quite a bit after the Fusion.
One that hath name thou can not otter
Also, think of what this means for laptops. First, you save a huge amount of space by not having to have a separate GPU chip on the board. Have you seen how crammed the mainboard is on the macbook? And with the significant improvements in power consumption, it's a win-win for the laptop market.
Intel is also leader in performance/watt, due to a complex power delivery architecture and better processor production facilities.
As long as you look only at raw CPU performance and power usage. Add GFX perf into consideration and...
(plus that would be quite recent development for Intel; their power consumption numbers weren't that great by themselves, when adding also chipsets of previous gen)
One that hath name thou can not otter
Looks like the Quickie Mart has a lawsuit on their hands.
There are no loopholes. It's either legal or it's not.
The 6-core Intel processor is the Extreme Edition (always was introduced at $1000)
If ((not realistic for server marker) && (cant sell for less than $1000 without undercutting our other offerings))
{
setlabel("Extreme Edition");
}
Where is Intel's budget 6-core design? Is it because they refuse to make budget 6-core CPU's, or is it because they can't make budget 6-core CPU's?
Either way, the proof is in the pudding. They are not targeting the highly parallel market either by choice ("ignoring that market" scenario) or by mistake ("caught with pants down" scenario)
"His name was James Damore."
Intel's on chip GPU is just that - a GPU, and a primitive one at that. It can't even do OpenCL. It's certainly not a competitor to anything that AMD will release. Never mind Intel's appalling graphics drivers (and consistent history of poor driver releases), and benchmark cheating (so that they look competitive in reviews).
AMD designed/implemented the 64bit instruction that will be running our desktop PCs for decades to come.
Intel was the one scrambling to catch up on that.
No sig today...
This demo was of Ontario - AMD's low power solution for netbooks and low-end notebooks. This will be using the low power Bobcat cores and probably something similar to an HD5450 graphics-wise.
I seriously doubt heat is going to be an issue.
hopefully this has good effects for cooling, also. Maybe genuises will stop designing boards with 2 hot components separated by 4-6" on a board cooled by 1 copper pipe/fan assembly ... cleverly heating everthing along the whole length of pipe.
Sounds like a non-advancement to me.
"Look, we can build a VCR *into* the TV, so they're in one unit!"
Yeah, so when either breaks, neither is usable.
Putting more points of failure into a device just doesn't sound like a great idea.
In the last 4 computers I've built/had, they've gone through at least 6-7 graphics cards and 5 processors. I can't remember a single one where they both failed simultaneously.
Now, if this tech will reduce the likelihood of CPU/GPU failures (which, IMO, are generally due to heat or less frequently power issues) somehow, then great. But I have a gut reaction against taking two really hot, power-intensive components and jamming them into even closer proximity.
Finally, I'm probably in the minority, but I prefer being able to take my components ala carte. There were many times in the past 25 years that I couldn't afford the best of all components TODAY, so I built a system with a very high-end mobo and CPU, but using my old soundboard, RAM, etc until I could afford individually to replace those components with peer-quality stuff.
-Styopa
It should be plenty good for space and power consumption. Just look at Intel's US15W chipset which includes the GMA500 IGP.
It's tiny and consumes 2W compared to previous gen chipset + GPU setups (GMA950) that consume 15W, lengthening the battery life by a huge margin.
The chip itself has good performance, hindered only by terrible outsourced drivers (Tungsten, I'm looking at you), currently only optimized for video decoding (who said two smooth 1080p streams at less than 100% CPU usage using EVR in MPC?)
Combining the CPU and GPU can probably give a comparable reduction of power consumption and size, with the support of AMD/ATi graphics core instead of PowerVR core + terrible Tungsten drivers.
^_^
This thing is going to smoke current CPUs in things like physic operations without the need of anything like CUDA and without the performance limit of the PCIe bus.
Ummm, but videocard has its own super-fast memory (and a lot of it), and it uses direct access to system RAM, while this little thing will have to share the memory access and caches with CPU.
without the need of anything like CUDA
I dare to say, that this is totally false.
AMD and Intel need to have a contest on the shittiest driver category. I have one of each. Each revision of xserver-xorg-video-intel bricks my laptop in a new and exciting way. And AMD's fglrx is a steaming pile of rendering errors, inconsistent performance, and crashes.
On the other hand, both Intel and AMD have released specs and participate in open source development. So in the long run, either one is a better choice than NVidia. So I'll continue to complain about them and submit bug reports. It's the open source way.
1^2=1; (-1)^2=1; 1^2=(-1)^2; 1=-1; 1=0.
Back in the days of Athlon64 vs Pentium 4 and Itanium AMD were ahead. Still since Core2 I'd say Intel are doing better. That being said Larrabee seems to be dead and I still think the idea has legs. Hopefully AMD will be to Larrabee what AMD64 was to IA64 - i.e. a more pragmatic version of the idea that ends up working better.
echo -e 'global _start\n _start:\n mov eax, 2\n int 80h\n jmp _start' > a.asm; nasm a.asm -f elf; ld a.o -o a;
cost - yeah, I can see Intel willingly passing the savings...anyway, cpu + mobo combo hasn't got cheaper at all
...)
This is where Intel's monopolistic behaviour rears its ugly head. In the past, the GPU needed to be integrated on the motherboard. Now it's on the CPU but Intel motherboard chipsets cost the same as previous generations. Seems like a terrific opportunity a market for 3rd party chipset vendors to make an offering (like the good old days when you could choose from VIA, Nvidia, SiS, Intel,
But wait, Intel will no longer allows 3rd parties to produce chipsets for their CPUs and keeps the profits from the artificially inflated chipset market to itself. Intel may have the performance crown, but its reasons like this (and the OEM slush funds to lock out AMD from Dell and other vendors) that keep me from supporting "Chipzilla"
From hell's heart I fstab at /dev/hdc
Uhmm, have any of you guys actually READ the tech specs for amd's 12 core parts? They're ALL QUAD CHANNEL DDR3 if G34 socket. Which given the pricing compared to Socket F means unless you had a fully equipped Socket F server (And didn't mind 8-core cpus max) you'd have no reason to stay with socket F and get just as much memory bandwidth as Intel's current top of the line processors.)
So while intel is currently giving more processing power per chip via hyperthreading, amd is giving just as much memory bandwidth at 1/2 to 1/3 the cost Go compare a 1567 Intel Xeon versus an AMD Opteron. You can have the whole mobo cpu combo for about the same as a single intel chip, which would get you 24 cores and a higher clockrate for less than you'd pay for 12 hyperthreaded cores via an 6 core intel chip. And that's with the same amount of quad channel memory bandwidth.
Now where things get interesting is deciding whether larger cache or cheaper price/higher clockrate is more useful for your application. The intel parts are available in 6 and 8 core parts with 18 and 24 megs of L3 cache, which the Opterons across the board are 12 meg. For many applications the performance penalty, if any, might be small, but for those that are cache hungry you could see a 5 fold(?) increase in performance compared to having to hit main memory.
But if it's multithreaded to begin with, you could also just throw a second server's worth of cores at it for the same price as the intel box :D
Yeah, right.
I'll believe that when I see it. For one thing, there's no way to expose that in any reasonable way to the OS. For another, that would mean the execution unit would have to answer to 2 contending schedulers...
Color me skeptical.
As opposed to how graphics drivers are security issue now?
Graphic cards can DMA memory and GPU can access pretty much any physical memory in the system (as long as it is visible via PCI bus). There is no simple fix for that but there are certain security features already available on graphics cards. Go read radeon Linux kernel sources, look at the command buffer parser (linux/drivers/gpu/drm/radeon/r600_cs.c for instance) that verifies that graphic card only accesses memory that belongs to it.
Also, there was some driver exploit in signed Windows graphics drivers that allowed loading unsigned code into windows kernel.
In the old days, there was a physical chipset which sat between the GPU and the CPU.
But in this architecture, there is no physical barrier - they're on the same silicon.
Look for the bad guys to try to force the graphics drivers to sneak over and sniff the memory of the CPUs - I can imagine how they might be able to load some code in a pr0n movie that could tell some pointer in a GPU driver to point to addresses of cache which [at least ostensibly] belong to a CPU, at which point they should be able to read the cache.
And if they're lucky, their specially-crafte pr0n-videos might even be able to WRITE to the CPU cache, at which point they can probably pwn the entire operating system.
Hopefully AMD has put some thought into their implementation, and has some sort of hardware safeguards that force the GPU to always act as the "slave" of its masters [the CPUs], but, if not, then all Hades could break loose.
[And Intel probably won't put nearly as much thought into their implementation as AMD did with theirs.]