Monday, May 2nd 2011
AMD Llano Fusion APU to Feature Radeon HD 6550 Graphics
AMD's upcoming Llano line of accelerated processing units (APUs), to compete with Intel's Sandy Bridge LGA1155 processors, is said to embed what AMD will refer to as the Radeon HD 6550 graphics core. Unlike Sandy Bridge where a processor die is simply fused with a integrated graphics northbridge onto a single die, Llano will see the GPU part of the silicon integrated with the rest of the APU in many other levels, including assisting the x86 cores with serial processing loads.
Llano's embedded GPU carries the AMD Radeon SKU of HD 6550. It will feature on AMD's Fusion A8-3550 and A8-3550P APUs, is DirectX 11 compliant, has 400 stream processors, and a core clock speed of 594 MHz. It uses memory shared from the main memory, but in all likelihood, AMD might work on SidePort-based memory support. Further, the Radeon HD 6550 can work in tandem with discrete AMD Radeon HD 6570 and HD 6670 "Turks" based graphics cards in the same way as AMD's IGPs have been known to work with entry-level Radeon GPUs using Hybrid CrossFireX. When the HD 6550 iGPU is working in tandem with HD 6670 or HD 6570, the graphics hardware will be recognized as "Radeon HD 6690".
Source:
DonanimHaber
Llano's embedded GPU carries the AMD Radeon SKU of HD 6550. It will feature on AMD's Fusion A8-3550 and A8-3550P APUs, is DirectX 11 compliant, has 400 stream processors, and a core clock speed of 594 MHz. It uses memory shared from the main memory, but in all likelihood, AMD might work on SidePort-based memory support. Further, the Radeon HD 6550 can work in tandem with discrete AMD Radeon HD 6570 and HD 6670 "Turks" based graphics cards in the same way as AMD's IGPs have been known to work with entry-level Radeon GPUs using Hybrid CrossFireX. When the HD 6550 iGPU is working in tandem with HD 6670 or HD 6570, the graphics hardware will be recognized as "Radeon HD 6690".
34 Comments on AMD Llano Fusion APU to Feature Radeon HD 6550 Graphics
Wouldn't make the most exceptional gaming laptop but I'm just thinkin about all those NV laptops where they used dual GPUs. Ya know, 9400M+9600M. Cept those weren't in SLI. You could only use 1 GPU at a time. But if it was done in a Crossfire setup where you could turn on both GPUs when performance was needed...kinda like that idea. I know power consumption and such might be an issue.
It's my main gaming rig, I have it connected to the receiver, tv, a projector and a pc screen 5 meeters away with a usb extension to a usb hub where I do other gaming.
I have 1 rig that rules them all! and I've tried to have a PS3 for loan for 4 months, a 360 wii, media center, nothing really beats the pc, it just does everything, htpc's can be rather powerfull, a 5570 is just not enough, a 5670 in terms of performance can turn into a htpc fusion chip.
Just look at my specs for my htpc, rather impressive, it impresses me to this date, and its a year old already.!
Link to HT RACK
its a GD04B from silverstone.
Link to GD04
its:
2x 5850Crossfire
6 core 1055T
8gb 1866 mhz (for new sandy or bulldozer)
80gb ssd
1tb samsung spinpoint
hx750.
in movies i can't hear it, the crunching from popcorn is far far far louder, get a good gpx cooling on the gphx and cpu cooler (8700cnps) with noctua case fans and your on!
as it works now, if you pair whatever it is, lets say.
HD6550 GDDR5 and HD6550 GDDR3, (if they exist) you would throttle down the GDDR5 to the GDDR3 speeds (effectively) in the crossfire setup.
as if you do 6950 1gb and 2gb in crossfire you only get 1gb memory, this is where the problem is!
On the other side, opencl, and gpgpu, osx uses this quite alot for such a young tech to be already, you bring opencl to so many people, and games can utilize opencl's capability of running tasks on a gpu.
No matter if you use the gpu or not for graphics, it can be used, I hope they are going to utilize this possibility, we only need intel to support it now!
and I hope all cpu's will have a on core graphics controller, no matter how slow it is, when a gpu dies you can still surf the webs, I see hardly any cons as long as they can utilize gpu's more effectively in applications.
display outputs should be usable, mine works (890GX) with HD5850 crossfire.
But in my post above you see limitations due to memory subsystem and how multigpu works with memory, unless lliano manages 28gb/sec as bulldozer will, I'll see very very memory bandwidth hungry 400 gpu cores. (hydra manages memory the smart way, but terrible scaling)
Apart from that, yes, this is definitely way more powerful than onboards. Socket FM1
www.amdzone.com/phpbb3/viewtopic.php?f=532&t=138344
imagine 400 cores;)
Limitations is that gpu's cant calculate all the things cpu's can, but when they can, they ROCK!
opencl, watch amd's video's, it involves the code, I know hardly any coding since I deal with servers, network, cisco cli, the closest is that i can make balls bounce in opengl and powershell scripting but I understood it.
dumbed down example.
opencl capable gpu present : use for this code.
if no opencl capable gpu present use cpu.
there, no requirement for gpu, can run on cpu, but will run on gpu if present, and all on 6 lines of code or something.
www.youtube.com/watch?v=ecYIsu83c0I
here's the links, its not AMD tech, it can be used by anyone, and I really hope it breaks through!
the video's are rather technical, but its worth a view if your interested.
But back to orginal subject again, for gaming this is a huge step forward, this design allows for smaller motherboards, removing northbrigde/igp alltogether, loosing a powerphase for it, and increasing performance for entry level computing while maintaining all the supports of the dedicated graphics, and the calculation power, the memory bandwidth for gaming is what worries me.
There is a very interesting article called "Debunking the 100X GPU vs. CPU myth: An evaluation of throughput computing on CPU and GPU" (Google it, you might need academic access to read it though) which has a comparison between a Core i7 960 and a GTX280 (the article is a bit old by now). The average computational advantage for the GPU is around 3 times as much, with the maximum shown in the article around x11 or x12 times faster. The problem shown within is the fact that many studies earlier compared highly-optimized GPGPU libraries with unoptimized CPU code. Once both are well optimized, the results aren't as amazing as GPGPU advocates may lead us to believe. This is not to say that there aren't applications where GPUs are indeed x100 times faster, but these remain as relatively isolated cases.
The bonus here will be some additional computational power for free, not a huge amount of computational power to eclipse the CPU. This is the right path, though, and better coupling between the CPU and GPU parts may well be the future.
there is tasks the gpu is actually slower, and therefore is per task seperation of gpu/cpu selection a good thing in opencl.
but again, we do not have compute power problems, a 1055t hardly gets loaded, but here i am with a 1055t with 2x5850CF, I either use alot of cpu, or half of cpu and lots of gpu power(games)
so if apps could use my graphics cards and replaceing two-three cores then you've gotten alot more performance per mm^2 and per watt, the same goes for the apu's
But I think none of this will be seen this year, next year but I start to see things supporting use of gpu compute power. most screens are 1080P, but for 1680x1050 I see no issues, as long as memory bandwidth is there.
Light gaming, on budget laptops, 1366x768 resolution, perfect
Sure, Crysis will have issues, and BF3, and Metro 2033, but even then you'd get away with it on lowered settings (contrary to what some PC hardware reviews suggest, games are still fun on low graphics settings :P).
One would *think* it could be connected through HyperTransport. My memory is fuzzy, but iirc it's spec is currently 32-bit/25.6gbps. That would mean that GDDR5 could be connected at up to 6400mhz to saturate it; 10% less than spec for 7Gbps and something very in tune with how AMD sets memory clocks. I don't know how well GDDR3/5 play together, but that opens up the possibility to play nice with discrete cards using cheap low-power (3200mhz) GDDR5 on a 64-bit or 1600mhz GDDR3 on a 128-bit bus (if bandwidth/timing parity is what is needed, and the '4x' clock of GDDR5 could be used in conjunction with the '2x' clock of GDDR3).
At any rate, with GDDR5 currently being 2Gb, and the power of the APU being as it is, no more than 2 chips would really be needed (512MB), although I could see the appeal for 4. Paired with a similar 512MB/1GB card like Redwood/Turks or even a small, low-power next-gen SI part, you could have a pretty bad-ass solution for resolutions that don't require 1Gb of VMEM (ie laptops or 'regular' non-1080p/+ consumer desktops) or perhaps even decent gaming at 1080p if they do end up going 1GB.
Going forward with higher density/speed and/or lower power RAM, more and/or faster HT links along with process improvements allowing greater GPU power within the APU, this kind of low-end crossfire could easily become a very formidable good-enough solution for the mass consumer...Even catapulting laptops out of their second-rate funk for gaming...at conceivably a low-cost to boot.
We shall see how it unfolds, but there is a truly exciting convergence/transformation of the PC platform emerging.