Archives Discussions

[rahulgarg]

fp64 and other extensions

Which device capabilities does Llano support? And specifically, is fp64 supported on the GPU? I am assuming not, but a confirmation will be nice.

[mikewolf_gkd]

fusion is only for notebook, its GPU core must be low, I think it will not support good opencl cability.

[Meteorhead]

Mike, fusion is not only for notebook. Sabine is for notebooks, Lynx platform is for desktops. Anyhow, it will most likely not feature DP. I once discussed this in a thread (with Micah if I'm not mistaken) and from rough speculations we concluded that adding DP capability porportionally increases the cost of smaller chips more than the larger ones. (I did not quite understand the argument)

Anyhow, feel free to add +1 to pseudo-DP in the Suggest feature you want in AMD APP, and then there will be DP support on it. 😉

[laobrasuca]

Roughly speaking (maybe very, very roughly), Llano is a Athlon/Phenon II up to x4 + Radeon up to HD6550 put together, so chances to have DP units in GPU are really small, not to say zero. But, since they are not for sell yet, don't expect to have AMD people commenting on this.

[Meteorhead]

Llano has been officially released yesterday, so I take it not much secret remains about them. All details have been specified and notebook manufacturers already announced their exact products based on Llano.

It is definately not Athlon. 1MB of L2 / core is even more than on Phenom II processors (512kB). Husky should outperform Phenom II by about 6% in terms of computation and has twice as much cache. That is quite far from an Athlon II.

[himanshu_gautam]

LLano APU does not have a DP supported GPU. More information will be released shortly.

[MicahVillmow]

LLano uses a GPU roughly equivalent to a Redwood XTX(A8), Pro(A6) and smaller(A4).

[Meteorhead]

May I ask why is it that AMD has mobile CPUs that are always one generation behind desktop variants? Llano is based on Phenom II (whatever the name was of the chip inside). Second generation APUs (showed off at Fusion Developer Summit just now) already has a working APU based on Bulldozer.

Since Bulldozer was supposed to launch alongside with Llano (meaning the chip design was ready), why not have both of them with the same technology? Intel Core-ix always uses same technology for desktop and mobile variants, and notebook processors are damned strong compared to AMD. I know APUs have the emphasis on the IGP, but why not use same architecture for the CPU part?

Also, IGP is one generation behind. Same will go for Trinity, which uses Bulldozer CPU (which will be 6 months old technology) and Cayman SIMD design IGP, which yet again will differ from HD7xxx series (most likely). The reason why I ask is, because Hybrid-CrossFire and the likes cannot be utilized with the same generation discrete GPUs, because they are always using a different architecture. Why not utilize all new designs at roughly the same time, so that products can co-operate better?

[MicahVillmow]

Meteorhead,
While that would be great to do, in practice it is not easy. The GPU can break backward compatibility every generation because really no-one directly targets the ISA except for the AMD shader compiler, the CPU cannot do that so just from a testing side of things, the CPU must make sure that pretty much every OS works without bugs. This involves lots and lots of testing that takes a lot of time. So, lets assume that to test every OS and all major applications takes a year of time(which isn't that far off as bulldozer was announced last august and announced to be released in Q3 `11), and by the time the CPU is released, the GPU is already on the next generation.

I think the real issue here is with the second half of your last paragraph. If the hybrid crossfire software worked in this situation better, there would be no issue, right? Since having different generations of GPU in a machine is an issue that will not go away in the future, AMD will have to improve its software, which we are doing.

[Meteorhead]

So if I take it correctly, GPUs can be released faster due to the reason that they must only comply with AMD shader compiler. In the case of CPU, it takes a lot more testing, thus when the chip is designed, by the time the testing is finished, the IGP falls one generation behind descrete variants.

In theory, isn't it possible to test the two seperately? It is not just software engineering that utilizes interface approach to modular development. Apart from a memory interface and perhaps a cache coherency protocol (if it is shared) the CPU and the GPU are still pretty much seperate inside an APU. Is it not possible to put them together if the interface is well defined?

So you say that the reason Intel IGPs are far inferior to AMD's is because they never create descrete variants, just integrated ones that fall behind the rivals due to the amount of testing? This statement however still does not answer why the desktop and mobile variants of CPUs have to differ on AMD side.

Two questions were: the IGP/discrete generation gap, and desktop/mobile VPU generation gap. If Llano and Bulldozer were to be released at the same time, why couldn't both utilize Bulldozer architecture? Since Trinity already has a working prototype, it seems to me that it really wouldn't have been so impossible to skip Llano. Or am I jumping to conclusions?

Having heard rumors about release dates of Trinity and Souther Island (7xxx) discrete cards, I would rather wait for that combination.

[MicahVillmow]

Meteorhead,
I don't know much about the CPU side, but modular development doesn't help when you have to do full product testing, even with well defined interfaces. It is this step alone that can make any CPU+GPU based product fall behind a GPU only based product. Even with all things in development and production being equal, the GPU doesn't have to be tested to run any OS's or any legacy applications. Sometimes there are hardware issues that would normally stop shipments on the CPU, but on the GPU the compiler might be able to work around them allowing us to ship hardware.

As for testing, I don't know must about the test infrastructure, but lets assume you want to test a brand new CPU. You aren't going to load up the latest Win7/Linux kernel and start running tests. You could do this, but these are extremely complicated pieces of software and it would be very difficult to track down issues that might occur. Instead you start on something simple like DOS or just the BIOS and run a large battery of tests there. Then as you get more confident in your hardware, you move onto more and more difficult tests. This takes time and on the GPU the amount of testing required for new hardware is miniscule in comparison.

I can't comment on Intel's IGP solution, but their technology is far behind AMD's when it comes to graphics. That is a design issue, not a testing issue. For the desktop vs mobile issue, you would need to ask that to one of the heads of development, I don't think any of the developers here would know the answer to that question.

Also, if AMD has a new product now, and a next generation product coming out in the future, why would they skip releasing the product now? That doesn't make any sense from a business perspective as you are just giving up revenue.