By using this site, you agree to our Privacy Policy and our Terms of Use. Close
Netyaroze on 02 June 2013
Edit Thread
Pemalite said:


It is just semantics. They both share the same memory pool, both the processors can talk to each other via that memory pool, it's been well documented in the linux community where they develop there own graphics drivers, the difference is GPU's of today are far more progammable and can talk directly via it's own dedicated bus to the CPU instead of passing through the System Ram like with Intels Ring Bus.
But that doesn't mean it's not possible for a CPU and IGP not to talk to each other and provide assistance in older platforms via slower methods. (Which is the point of this entire discussion.)
http://en.wikipedia.org/wiki/Cache_coherence
https://01.org/linuxgraphics/blogs/vivijim/2012/i915/gem-crashcourse-daniel-vetter
http://supercomputingblog.com/cuda/cuda-memory-and-cache-architecture/

That link you provided is also not in English, no idea why you would post it in another language when this entire conversation has been purely in English.
Here is the English version for those wanting to know: http://en.wikipedia.org/wiki/Non-Uniform_Memory_Access
Again, Intel had supported NUMA 5-6 years ago, starting with the Nahalem platform.
http://lse.sourceforge.net/numa/faq/

Intel will be taking it a step farther with Haswell with it's embedded DRAM that both the CPU and IGP utilise as a cache and can see what processor has what, much akin to what they do in System Ram, it is just another step up in the cache hierachy to hide the latency and bandwidth deficit that DDR3 has. (Only on the GT3e IGP's however.)

Now, I honestly don't see dedicated memory going away, how you came to that conclusion beats me.
For example, back on the Radeon Xpress/3000/4000 IGP's AMD included 32Mb - 256Mb of DDR3 (Or better) Ram dedicated to graphics. - http://www.anandtech.com/show/1537/17
Intel has gone with throwing a bunch of transisters for a large fast cache on the CPU die.
http://www.anandtech.com/show/6911/intels-return-to-the-dram-business-haswell-gt3e-to-integrate-128mb-edram
AMD will be introducing GDDR5 for Integrated graphics soon.
http://www.xbitlabs.com/news/cpu/display/20130305060258_AMD_s_Fusion_Kaveri_APU_Supports_GDDR5_Memory_Report.html


Lastly... As for PCI-E 3.0 bandwidth... It is 1GB/sec per lane in each direction, maximum of 32GB/s in total, PCI-E 4.0 will double that again.

 

Sorry for copy/pasting the wrong link I thought I copied the english version.

 

I am not sure we are talking about the same thing.

I never doubted this: "But that doesn't mean it's not possible for a CPU and IGP not to talk to each other and provide assistance in older platforms via slower methods. (Which is the point of this entire discussion.)"

This was never my point. I dont see how you took this from what I wrote.

 

I wrote: "I am refering to unified Ram which is just way more efficent and will come for PC in the future, also CPU GPU communication is better so in a sense they are ahead"

 

(I should've specified it that I dont mean just any normal unified RAM.)

 

You wrote: "PC had unified memory systems for decades, what do you think APU's are? What about any system with an IGP? The memory isn't split on such systems, the PC had unified memory systems far before the consoles, infact it was invented on the PC first."

 

Yes PC had unified memory systems but not HUMA thats different than what Intel did. It needs way less copying around data from pool to pool which uses CPU/GPU time, ram bandwith/space.

 

http://www.bjorn3d.com/2013/04/amd-huma/#.UarVu9jVb2s

"With hUMA, all processing cores (CPU and GPU, etc) share single memory address and operates at the same bandwidth. CPU and GPU caches both see the data stored in the memory and can access and allocate any location in the system’s virtual memory. Data no longer needs to be copied back and forth between the CPU and the GPU before they can be seen. The CPU can pass a pointer and its entire data structure to the GPU. The result from the GPU can be read directly by the CPU without the need to copy."


If its just semantics WHY are people excited (google it) why is anyone even talking about it. And how has this been done years ago ?

 

 

That dedicated Ram goes away is just a guess and the following just a personal opinion: I dont see why sending data between two pools of ram and CPU/GPU is benefitial for anything. I also think that HSA/HUMA is a natural progression. I think the unification of GPU and CPU is under way for years. And the design philosophy behind the consoles is what AMD/Intel ultimatively are aiming to achieve, on a broad scale on the PC market. This is obviously just a guess. But getting rid of GPUs and CPUs seems to be a long going trend. Intel with his Larrabee tried it before. Cell is the same basic idea. Now APUs. And using two pools is pointless anyway. I can see a (distant) future for Desktop PCs were you can just have a board with 1 chip and 1 Memory.  And if more power is necessary use two boards.

 

PC obviously will have the technology, but not on the same scale yet.

 

There are some things were Consoles are up to date and can expand on concepts due to not being tied to the modular PC Ecosystem. That is my whole point. Always was.

If you still think I am wrong please send me your answer per PM. I dont want to derail this thread anymore, if people see a problem with it.